[Edit:
Version v5.0 works with latest comfyui (v0.15.0).
If you have any problems, please refer to the FAQ at the bottom of the page or have a look in the comments.
Many thanks to everyone who tested this workflow. Thank you very much for the many inquiries and, of course, for all the knowledge and experience you have contributed. here👍🙂
Special thanks to:
@SeoulSeeker for the "Dead Simple MMAudio" workflow wich are the basis of the audio part here,
@taek75799 for the really well working enhanced models
@Bakazaya pointing to the color issue in version v3.0 and running lots of tests,
@bluntfeather sharing latest experiances with installing Comfyui-Easy-Install,
@nitrovtx for remain persistent in matters of quality and running a lot of tests,
@Icey64 for providing the link to "Comfyui-Easy Install",
@boinobin730 for asking for a First to Last Frame option, running pre tests and responding fast as hell 🙂 and
@SnowShoes311 thank you so much again for all your buzzing 😋]
Features:
Optimized Wan 2.2 workflow, runs perfect on RTX 3060 12 GB VRAM GPU and 32 GB RAM,
"Text to Video", "Image to Video" and "First/Last Frame 2 Video" generation in one workflow, all with easy audio generation,
easy installation/model downloading, all necessary sources are specified,
easy to use workflow, clearly structured, all necessary steps are explained,
easy switches for mode selection,
easy prompt selection for fast prompt creation/testing,
easy switching between "standard" and "enhanced" models,
very fast and smoth high quality outputs up to aprox. 1440 x 960 with 60fps,
2x fast upscaler,
4x fast framerate multiplier,
MMAudio Sampler (generates sound accordingly to the video action),
Triton and Sage Attention option,
A 5 Second long high quality video generation takes about 10 - 15 minutes (see below).
Tested generation times:
As a rough guide value for RTX 3060 GPU: generating a 5 second long high quality 1440 x 960 60 fps video with 6 steps it will take:
t2v: around 10 - 12 minutes,
i2v: around 15 minutes.
Comfyui-Easy-Install with Triton + SageAttention:
This workflow should work with any latest comfyui version >v0.6.0 (Desktop, Embedded, Windows/Linux).
However, comfyui is developing rapidly, and it often happens that some of the custom nodes used are not updated quickly enough or not updated at all. Manual workarounds are sometimes necessary. Furthermore, care must be taken to ensure that there are no conflicts with other nodes.
If you're having difficulties with your existing comfyui system or if you want to run video generation on a separate (parallel) comfyui system, like I do, I would recommend you the following installer: https://github.com/Tavris1/ComfyUI-Easy-Install.
Complete installation of comfyui including manager and some pre configured custom nodes is just one click - really 🙂
Installation of Triton + SageAttention is just a second click - really 🙂 And since it's so easy now, I would definitely recommend it to you for video generation.
Cause it is an embedded version, you can install it parallel to your existing comfyui version without the risk to ruin your working system.
After installation just configure the "extra_model_paths.yaml" file to use your existing models.
After a fresh installation of Comfyui-Easy-Install you might have some issues too, but there are known workarounds - please see the FAQ below.
For testing/understanding/experimenting/changing the workflow:
Click "Toggle Link Visibility" to see the links.
click the Subgraph symbols to open the Subgraphs.
for quick testing you may lower the settings for: steps, clip lenght and video resolution,
be really carefull with modifying Groups or Subgroups (even Titel or Color) cause they are essential for switching,
feel free to try and test other models. Just give me a hint if you find models which deliver better results and fitting the 12 GB VRAM limit.
And as usual: Have Fun 🙂🙂
Short Conclusion:
This workflow is based on elements of a variety of allready published workflows. My "job" was only to put things together, optimize it for a small machine and create a most simple and hopfully user or even "beginner" friendly workflow.
I`m not an "expert" - just a user who wants to get it running on "available" hardware.
There are many things I don't really understand. If you find mistakes or better solutions please give me a hint.
And I really hope that even "beginners" have a chance to go the first steps...
Frequently Asked Questions (FAQ):
For quick and better overview I will try to merge all known issues here - step by step (please be patiant). If your issue is not listed here, please have a look in the comments first. Most issues have been allready discussed.
Comfyui Nodes 2.0:
Turn off Nodes 2.0 in comfyui (use comfyui menue). Actually not all custom nodes are supported.
Comfyui crashes after generation while vae decode, upscaling or frame rate multiplying (Rife VFI) without any error report:
This is a RAM problem (not VRAM). Increase your swap file (min. 64 to 128 GB) or set it to automatic management on a fast drive with at least 100 GB free space.
JW Nodes (JWFloatToInteger, JWIntergerDiv, JWImageResizeByLongerSide), soundfile missing:
For the workaround look here and here:
python -m pip install soundfileFresh Comfyui-Easy_Install Installation (missing soundfile and Pytorch v2.9.0 issue with SageAttention on Windows:
For full conversation look here.
Open cmd in python_embedded folder:
python -m pip install soundfile python -m pip uninstall -y torch torchvision torchaudiopython -m pip install torch==2.8.0 torchvision==0.23.0 torchaudio==2.8.0 --index-url https://download.pytorch.org/whl/cu126Slider Nodes - how can I modify the "default" values:
Right click the slider node, choose Properties and set the values you like 🙂🙃
Description
Fixed:
All 4 input fields for prompts working now for T2V too.
FAQ
Comments (322)
UnetLoaderGGUF
expected str, bytes or os.PathLike object, not NoneType
LoraLoaderModelOnly
Model in folder 'loras' with filename '91 Lighning\Wan21_I2V_14B_lightx2v_cfg_step_distill_lora_rank64.safetensors' not found.
The contrast fades out a bit after the first half second or so.
So you were able to solve your issues yourself?
@arkinson No, it's still doing it. Any ideas?
@camarcuson194 No, I mean your other two comments. Once you have found a solution, please explain on every comment what you did so that others can benefit from it too.
Ok, to your last question: I experience the same behaviour on my generations somtimes too. One idea would be to experiment with the lightx2v Lora: disabling/changing weights or testing the newest Wan 2.2 Loras. You also meight have a look here: https://civitai.com/models/1852904/wan-22-workflow-optimized-for-rtx-3060-12-gb-vram-gpu?dialog=commentThread&commentId=955554
@arkinson No clues on those either. I'll take a look at the reference, thanks.
Thanks for this! I had so much fun with 2.1. Alas, I get this error when upgrading to your 2.2 version when it reaches the Ksampler High node:
!!! Exception during processing !!! 'NoneType' object is not subscriptable
Traceback (most recent call last):
File "D:\StabilityMatrix-win-x64\Data\Packages\ComfyUI\execution.py", line 496, in execute
output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, hidden_inputs=hidden_inputs)
File "D:\StabilityMatrix-win-x64\Data\Packages\ComfyUI\execution.py", line 315, in get_output_data
return_values = await asyncmap_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, hidden_inputs=hidden_inputs)
File "D:\StabilityMatrix-win-x64\Data\Packages\ComfyUI\execution.py", line 289, in asyncmap_node_over_list
await process_inputs(input_dict, i)
File "D:\StabilityMatrix-win-x64\Data\Packages\ComfyUI\execution.py", line 277, in process_inputs
result = f(**inputs)
File "D:\StabilityMatrix-win-x64\Data\Packages\ComfyUI\nodes.py", line 1555, in sample
return common_ksampler(model, noise_seed, steps, cfg, sampler_name, scheduler, positive, negative, latent_image, denoise=denoise, disable_noise=disable_noise, start_step=start_at_step, last_step=end_at_step, force_full_denoise=force_full_denoise)
File "D:\StabilityMatrix-win-x64\Data\Packages\ComfyUI\nodes.py", line 1473, in common_ksampler
latent_image = latent["samples"]
TypeError: 'NoneType' object is not subscriptable
@jandroplul Hi - thank you. "TypeError: 'NoneType' object is not subscriptable" just means there is some input missing. Please check you really have selected YOUR local models in the model loader nodes.
@arkinson Thanks for the suggestion! I found it, I missed 1 of the switch nodes. Now it's working! yeey!
@jandroplul I am glad you got it running 🙂
Thanks a lot for this workflow is so simple and well explained, works perfect on my 5070ti
Hi - thank you so much for your feedback. Happy generating 🙂
Can you make the option to turn on and off upscale. Sometime we need to test if the result is good before we want to scale it.
review : It take 6 minutes for 3 second video and 12 minutes for 5 second video.
@youtube02 Hi, the upscaler works very fast, but you are right, the multiplier needs some time.
You simply can use the preview video to decide if you like to wait for upscale/multiply or not. If not, just cancel the generation process and start a new one.
Basically I would not like to make the workflow more complicated than necessary. If you really need this function, maybe for automatic batch preview generations, just disable the nodes or move the nodes to a new group and use the rgthree Fast Groups Bypasser: rgthree/rgthree-comfy: Making ComfyUI more comfortable!
Hello, first of all thank you for this workflow. I have one question tho. https://civitai.com/posts/23833666 How do i make it so that my videos arent that blurry? Thank you. I am a complete beginner and most guides don't have a section on blurry videos.
Nvm, found it out. The lighting loras were at a value of 3. Just had to change that. And once again this is amazing thank you!
@Erb Hi thank you so much and I´m glad you got it running 🙂
To the lightx2v Lora: I use the old Wan2.1 Lora in the workflow. My default weight settings (High = 3.0, Low = 1.5) should work in most cases.
But you are right, in some cases it produces contrastless or even blurry videos. Adjusting the weights might help.
If you like to experiment you can also try the latest Wan2.2 Loras and you might have a look here: https://civitai.com/models/1852904/wan-22-workflow-optimized-for-rtx-3060-12-gb-vram-gpu?dialog=commentThread&commentId=955554. (look in the middle of the discussion at boinobin730`s comment for the right download link).
@Erb @arkinson What values did you give the lighting loras?
Edit: I lowered the Lighting lora values to 1 and 0.5 and the videos are coming out great! Thanks!
@Maximus__Prime Thank you for your feedback 👍 Never tought it will work with such low values 🙄 I accidentally run a few tests with 1.0 and 1.0 some times ago and it had produce a nice "vintage" look. Seems I have to do more investigations for myself now🙂
@arkinson What values do you typically use? Thanks for your help!
@Maximus__Prime With the default Wan2.1 lightx2v Lora: High = 3.0 and Low = 1.5 and with the Wan2.2 lightx2v Lora: 1.0 and 1.0.
If you lower the weights for the Wan2.1 Lora to 1.0 and 1.0 for example it seems to reduce the contrast dramatically.
I did some more test runs with the latest Wan2.2 Lora again - it definately produces better "quality" (more contrast and brightness) but it is harder to get movement. Finally i swiched back to the "default" Wan2.1 Lora.
By the way - trying to help others I learnd a lot for myself too and mostly I get positive and usefull feedback 🙂
@atrociraptor Hi -thank you so much for buzzing 😋🙂
And thank you for publishing this workflow, it's great!!
@atrociraptor Happy generating 🙂
@arkinson I've a question, if you don't mind? I saw people using a checkpoint when generating their videos. How does one do that?
@atrociraptor You mean the checkpoints/loras mentioned in civitai under "Resources used"? You have to add the metadata menually in civitai.
@arkinson Oh apologies, I meant what node would I use in this workflow and where would it be placed? If it's too much of a bother, no worries!
@atrociraptor Sorry, but I don`t got your question. What would you like to do exactly?
@arkinson Is it possible to add a node that loads a checkpoint in this workflow? For example, this one: https://civitai.com/models/1713721/ratatoskr-wan-21-wan-22-hybrid-i2v-animal-creature-and-furry?modelVersionId=2074025
@atrociraptor Ahh - I got it. Short anwer: To load a checkpoint model use a Checkpoint Loader node instead of the Unet Loader nodes.
Long answer: My workflow is completely optimized to run with 12 GB VRAM, especially the models I selected after a lot of try and error testing. If you like to use other models you have to do all this by yourself again and belief me - you might have good luck and find a way soon, or you can spent hours/days or weeks with testing a lot of stuff. The model in your link for example seems to have a lightning Lora included. Probably the best way for this model would be to start with a complete new very simple workflow to test the basics.....
@arkinson Gotcha! Thank you very much! :)
Absolutely love this!!!! Getting way better results.
@adefoss532 Hi - thank you so much and have fun 🙂
It has a habit of reverting back to the original image to end the video. I'm not using any end frame. Any idea on how to prevent this? Some sources say setting 81 frames will fix it, but I can't figure out where to actually set how many frames it is.
@NeilHamburger Hi - I see this behavior often too. In my experience, it depents from several parameters: start image, prompting, lightx2v Lora and of course the count of total frames (clip length). Experimenting with the clip length may be usefull. For easy use I "designed" the workflow to input the Clip Length instead of frames. But calculation is easy:
frame_rate * clip_length + 1 = total_frames
As you can see from the "Video Combine Preview" node our frame_rate for generating is 16 frames per second. Just set the Clip Length in the workflow to 5 seconds and you get 81 frames:
16 * 5 +1 = 81
@arkinson Thanks. I must be going nuts because I don't see any way of changing the clip length in the workslow. I see "Clip Lenght ( in seconds )" but it's just a reddish box but nothing else in it. I wonder if this is a Linux GUI bug or something. Maybe I'll hack up the json itself directly instead.
Edit: ah I think it's a ComfyUI bug with sliders. https://github.com/Smirnov75/ComfyUI-mxToolkit/issues/28#issuecomment-2603091317
Editting the json worked in any case.
@NeilHamburger Thank you for your feedback and the link for the workaround to bypass the mixlab conflict 👍
@vampirox22 Hi - thank you so much for buzzing too 😋😋🙂
Not sure why the requirements are never fully included in workflows, but you'll need many things for this. In the python venv (or general python installation, depending on how you installed it), you'll need these modules installed: yt-dlp, openexr-python, soundfile, opencv-python, gguf, and imageio_ffmpeg.
Open a terminal window in "ComfyUI/custom_nodes", or run a script from there, and git clone all of the following:
git clone https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite.git
git clone https://github.com/Fannovel16/ComfyUI-Frame-Interpolation.git
git clone https://github.com/pythongosssss/ComfyUI-Custom-Scripts.git
git clone https://github.com/city96/ComfyUI-GGUF.git
git clone https://github.com/orssorbit/ComfyUI-wanBlockswap.git
git clone https://github.com/kijai/ComfyUI-KJNodes.git
git clone https://github.com/jamesWalker55/comfyui-various.git
git clone https://github.com/DoctorDiffusion/ComfyUI-MediaMixer.git
git clone https://github.com/Smirnov75/comfyui-mxtoolkit.git
git clone https://github.com/bash-j/mikey_nodes.git
git clone https://github.com/rgthree/rgthree-comfy.git
Also, the lora link is wrong. You can find the required loras here: https://huggingface.co/chococka/wanloras/tree/main/2.2
Hope this helps someone.
@civai3kc64801 Hi, wich OS and wich comfyui version do you use? I have the comfyui Windows Desktop version and a comfyui embedded running. For both there is no need to install these moduls manually. Strangely enough no user has ever asked for this before 🙄
Lora link: I deliberately use the old Wan2.1 lightning Lora as default in my workflow and mentioned the newer Wan2.2 Loras as an alternative. Because, in my experience, the old Lora delivers the best results for most users. For the discussion about it, please look here: https://civitai.com/models/1852904/wan-22-workflow-optimized-for-rtx-3060-12-gb-vram-gpu?dialog=commentThread&commentId=955554
By the way, the new Wan2.2 I2V Loras are just 6 days old. Did you run allready some tests versus the old Wan2.1 Lora? If you have the capacity I would be glad about some first results.
Hi, the workflow works perfect, but I have one question If I want to start with higher resolution than 720 x 480, like 1000 x 1000 which parameters I must adjust? This start resolution has any reasson to be like this or is optimized for 3060? Im using 4070. TY
Hi - thank you. My suggestion for max. resolution is optimized for RTX 3060 only. With more VRAM you can try higher resolutions of course. For I2V you have to edit the max value in the slider node: right click -> Properties -> max.
Keep in mind: the used VRAM depents from a couple of parameters - generally: resolution, generation steps, clip length (total frames), frame rate, etc.
I've experienced a fair number of crashes -- it says "Killed" or something in the console and ComfyUI goes dead -- during the RIFE VFI phase. Memory issue, I'm guessing. Any ideas on how to make that phase less demanding?
@NeilHamburger I would guess a RAM (not VRAM) problem. Probably your swap file size is to low. Please have a look here: https://civitai.com/models/1852904/wan-22-workflow-optimized-for-rtx-3060-12-gb-vram-gpu?dialog=commentThread&commentId=957940
@arkinson Thanks. I figured 10 GB swap would have been large enough. Guess not.
@NeilHamburger Is it working now? What size did you set it to?
@arkinson I went up to 20GB and haven't had any crashes since. Usually around 650x650 resolution images.
@NeilHamburger Thank you for confirming it 👍
Hi - I would guess you have a wrong model selected, maybe the vae model. Please look here for example: RuntimeError: Given groups=1, weight of size [5120, 36, 1, 2, 2], expected input[1, 32, 16, 116, 90] to have 36 channels, but got 32 channels instead · Issue #9596 · comfyanonymous/ComfyUI.
Start simple: use at first T2V only and check twice you definitely have downloaded and selected the right models.
I wanted to say I really enjoy this workflow, I'm new to this and this is the first one I've been able to get working. The instructions within the workflow were extremely helpful in getting everything set up and working.
My only question is about the multiple text prompts - are they meant to be used together to create one video? different videos in a batch? or just as a means to have different prompts typed up and easily switch between them (I'm leaning toward this one now that I type this)?
@dragon509127 Hi - thank you so much for your feedback and your buzzing too 😋🙂and I`m glad you got it running 👍
Multiple Prompts: Your last guess is right. They are just for conviniance: You can use one text field for generating and the other ones for creating new prompts meanwhile. Or you can "save" some good "standard" prompts and easy switch between them, for example.
Please make sure you use my last workflow version v2.2 and you have selected only one option in the black prompts swich node.
Thanks a lot for the great workflow, I got it working on the first try. But, сan anyone tell me why when I upload the original image, the output is completely different, as if it relies solely on the hint, without using the original image. Everything was done based on your instructions and the correct models were loaded, and one switch was selected from the black switches.
@Jio_R Hi - thank you for your feedback 🙂
The first video frames have allways to be identically to your start image - so I guess you did not selected the "right" image. Please check this:
1. "Enable 02 Image to Video" yes
2. "Enable 01 Image to Video" yes
3. "Enable 01 Single Image" yes
4. "Input First Frame Image" input YOUR local image pass or drag n drop YOUR local image here
This should work out of the box.
@arkinson Thank you so much for your quick response and help. I looked at your original schematic again and realized my mistake: I selected T2V models instead of I2V everywhere. :) Everything worked! Thank you! P.S. Could you tell me how to add a voice, text, or audio file for lip-syncing to this schematic?
@Jio_R I´m glad you got it running. Sorry, I have no experiances with Wan + lip-syncing. Is there something on the "market" already, especially for Wan2.2? And how does that work?
@arkinson Well, at the moment it seems like InfiniteTalk, I even have a ready-made scheme, but I still can’t run it, it says there is not enough memory, my RTX 2060 12Gb did not cope with this scheme, so I thought maybe with your scheme and the WanVideoBlockSwap Low Pass block, maybe you could somehow get it running.
@Jio_R I have heard about InfiniteTalk, but I´m not shure if this works with Wan2.2 allready. If you publish a link and some desription to your workflow I might have a look at it, but please understand - lip syncing is actually not my priority. You have to test the basics by yourself.
BlockSwap: Just try the wan blockswap node in your workflow and test if it will work in combination with the models you use.
@arkinson I'm not very good at knowing how to connect everything correctly, as I'm new to Comfiui. I sent you a diagram in a private message (it's not my diagram; I'm just a newbie). I know this isn't your profile, but maybe you could point me in the right direction. If you have the time and inclination, let me know in a private message if you don't.
@Jio_R Please just provide a link here, cause I do not use civtai`s private massage system.
Just add the blockswap node after your model loader node.
@arkinson Thanks for the advice, I'll give it a try. Thanks again for your answers and advice, and for the workflow itself.
@bungel Hi - thank you so much for your feedback and your buzzing 😋🙂
Great work as always! If I might ask for the next version to implement TripleKSampler node so we can benefit from the first pass without lightning Loras. It really helps with motion. https://github.com/VraethrDalkr/ComfyUI-TripleKSampler
I'm going to take a shot at replacing the KSampler nodes in your latest workflow with the custom node, will let you know if I can get it working. Thanks!
@ainewb14 I started a new comment just to keep track. Please look here: https://civitai.com/models/1852904/wan-22-workflow-optimized-for-rtx-3060-12-gb-vram-gpu?dialog=commentThread&commentId=999729
@ainewb14 TripleKSampler: https://github.com/VraethrDalkr/ComfyUI-TripleKSampler
Hi - thank you for the idea with the TripleKSampler and the link. I just did a quick test with the t2v_simple.json workflow from the examples. I used the same conditions as in my workflow tests (resolution: 720 x 480 and 81 frames) and this is the result for generation time:
my workflow: max 12 Minutes (with upscaling and framerate multiplying),
t2v_simple: around 22 Minutes just for generating (without upscaling and multiplying). [edit: around 17 minutes after models already loaded]
Some times ago I already did some tests. I just used 3 KSampler. The problem is, you definitely have to run more then 6 steps (t2v_simple seems to use 10 steps). And ultimately, I was unable to detect any significantly better movements.
I would say this is a "no go" for small gpu`s. Did you run some more "serious" tests already? Wich gpu do you use?
@ainewb14 In the meantime I did another quick test with the i2v_simple workflow and SageAttention + Triton for a 8 second video. Ok - output quality looks good. Curiously these workflows using completely different lightx2v Loras and my machine was hard at the limit even with blockswap. And I got a lot of log errors, even these Loras seems to be Wan2.2 Loras. I will test these Loras with my workflow too next time.
If you or some others are interested and have capacities for more serious testing/experimenting please let me know and please share your results here. If you are interested I can share my extended test workflows (with Triton + SageAttention and BlockSwap) based on the example workflows.
@ainewb14 Oh my - seems I talk with myself 🙄 OK, after some more tests I can`t see any advantages of the TripleKSampler for my use cases. Sometimes it may help a little bit, but mostly I still get equal or better results with my existing workflow and the old Wan2.1 lighning Lora. For this reason, and in view of the general enthusiasm for this topic🙄😉 I will conclude here.
Thank you so much for this workflow. It works great and is really easy to understand and get working. The notes and download links/install locations are really helpful. I'm running this on an RTX 2080TI and have found it takes around 10 minutes to generate a 3 second clip with T2V and around 25 minutes for a 5 second clip.
I just had one question. I haven't been able to get the I2V Last Frame to work. It generates a video fine but it seems to ignore the last frame. Is the last frame of the generated video supposed to look identical to the input last frame? I've tried both videos of length 3 and 5 seconds and neither worked.
Thanks!
@StringCheeseTheory Hi - thank you so much for your feedback and I`m glad you got it working 🙂
1. 25 Minutes for a 5 second T2V clip?? Did you use higher resolutions? I have no experiances with your gpu, but I read the 2080TI shoud be faster then the 3060 - or am I wrong?
2. I2V First to Last Frame: The last frame of the video has to be identical to your last frame image. I just did a short test. My last version v2.2 works well on my side. Your issue is strange. Please check the swich nodes:
"Enable 02 First/Last Frame to Image" = yes
"Enable 01 Single Image" = yes
and check in both image loader nodes that you have selected YOUR local images.
And check too, you don`t have deleted any connections or nodes accidentally - or just redownload the workflow for a clean new test.
@arkinson 1. I just used the default settings of generating a 1440*960p video at 60fps. I'm not sure how exactly 2080TIs compare to 3060s but my GPU is also pretty cooked lol, it's getting a bit old and it's been through a lot. The runtime is in line with other workflows I've used except yours is a bit faster because it also includes the upscaling.
2. So I just tried again and now it is working. The only thing I changed was putting the First/Last images into the ComfyUI "input" folder and that seems to have fixed things. It could have also just required a restart or something. Either way now everything is working, thanks!
@StringCheeseTheory Uhh - so the RTX 2080TI takes about twice as long as the RTX 3060. In this case you might consider to install Triton + SageAttention. On the other hand you will get good quality outputs too if you reduce the start resolution to about 640 x 420 for example or even less.
For the first/last images is no special path required. Just drag and drop your images from any local path into the nodes.
@ChrissyS Thank you for buzzing 🙂
@Jio_R Thank you so much for buzzing 😋🙂
You're always welcome, I look forward to hearing about your new workflows! I wish you great creative ideas.
Hello, yesterday the setup worked like a charm, but today it won't start. It only reaches KSampler High and that's it. ComfiUI didn't update the nodes either, and the line hangs: Initializing block swap: 0%| | 0/40 [00:00<?, ?it/s] Can you tell me what the problem might be? I've tried it on different versions of ComfiUI.
P.S. I updated the nodes in the workflow and everything started working again. Maybe it was something in ComfiUI itself.
@Jio_R Any error messages? Cause it hangs while blockswapping I woud guess a RAM (not VRAM) problem / too low swap file size. Please look here: https://civitai.com/models/1852904/wan-22-workflow-optimized-for-rtx-3060-12-gb-vram-gpu?dialog=commentThread&commentId=992196
@arkinson As I wrote earlier, I simply updated the nodes of your workflow and everything worked again!
@Jio_R Sometimes life is easy as hell 😅🙂
@arkinson Well spotted!
Thank you so much for this.
This is my first time with comfyui and it works for me but (I think) I was a bit mislead by some of the instructions at the top regarding the subdirectories where to place files. The wan models placed inside of the "models/unet" folder were not found until I eventually moved them to "models/diffusion_models".
Then the main Lora was also not found because the workflow expects it to be inside "loras/91 Lighning/" instead of just "loras/".
Just sharing in case you want to update the documentation at the top.
Anyhow, thanks again!
@outsider21 Hi - thank you for your feedback. But I´m not sure if you got it really running.
In my instructions I give the hint that you have to click in every loader node and select the model from "YOUR" loacal path. It seems you don`t have done it, cause you try to load the Lora models from "MY" local pathes.
My instructions regarding the directory structure are correct als long as you use a standard comfyui version and you don`t have configured other/(wrong) subdirectories in your "extra_model_paths.yaml file".
By the way: Have you installed comfyui via "StabillityMatrix" for example? This "manager" software don`t uses the standard comfyui directories (heaven knows why) and indeed - you have to save the unet models in "models/diffusion_models" instead.
@arkinson Thanks for that feedback. It does work (not too sure how, though :D). I use the Windows app and when I click on some of the nodes, such as the Unet Loader Node to change the checkpoint, it doesn't do anything, the interface doesn't react, so I have to edit values in the json file of the workflow in order to make changes to things sometimes (like dropdowns). It's not ideal but it works 😅
@outsider21 Very strange. Seems there is something completely wrong with your system 🙄 What do you mean with windows app? Windows insataller for comfyui python version/embeddet version or comfyui Desktop version?? Latest update? Any errors while updating, etc. ??
@arkinson Don't worry about it, I've tried the git clone method where you launch the app via the terminal running main.py, and also the windows desktop app for ComfyUI. I'm sticking to the desktop app cause it's faster to install missing nodes. But yeah it has this issue with the UI for me.
Anyway, I'm more focused on finding out how to make it so that the animation loops seamlessly (starting frame and ending frame are the same). The default workflow doesn't seem to allow for this out of the box, but it's helping me learn the behavior of nodes :)
@outsider21 I use windows Desktop version too. There is definetely something wrong with your installation. Most issues are conflicting custom nodes. Have a look in the comfyui manager. Best practise is to delete conflicting node packs step by step and see if it work. But if you allready have edited json files I would really recommend you a clean fresh installation of the desktop version.
Simply use the first/last frame option of my workflow. This will do exactly what you want. And with the saved last frame image you can start a new generation process to extend your first video.
Let me ask, would you be able to create a purely Image to Video workflow? I'd be very grateful!
@Jio_R You can simply do this by yourself. Just reduce my workflow to the basics you like:
Just delete all the switch options with the "Any Switch" nodes and all parts you don`t need.
Or start with an official template and use my workflow to add what you want.
Or quick and easy: have a look here: https://civitai.com/articles/19794/closed-test-workflow Use attatched version v1. This is a simple quick and dirty I2V first/last frame workflow for testing. Simply replace the "WanFirstLastFrameToVideo" node by "WanImageToVideo" node and you get what you want 🙂
@arkinson As always, thank you for your prompt response and advice!
Thanks for the workflow, it was working 10/10 but now it shows Reconnecting when going through WanVideoBlock Swap Low Pass ( 5 sec 720 size), any guess? Also any chance to add more Lora's?
@mariogonzd813 Just look here: https://civitai.com/models/1852904?commentId=1002835&dialog=commentThread
You can add any Lora you want, see "Additional Loras" in the workflow.
@arkinson Thanks for your reply, unfortunately keeps crashing on Low Pass, even updating nodes. Guessing it is something to do with version ComfyUI Desktop is outdated and some nodes would require to update.
@mariogonzd813 Ok, the prerequisite is, of course, that your comfyui and your nodes are up to date. Have you checked your swap file, as I menttioned in the link???
@arkinson Yes sir, 30 GB now. Thing is yaterday was working flawless. And love the workflow to be honest, super complete. I think some node got updated and Desktop current version is the problem here. Tomorrow I will clean install Comfy with github latest version.
@mariogonzd813 I use the Desktop version too and have no issues on my side. Any error messages in the log files?
how to use custom model, do I still need model that only has HIGH and LOW? I'm about to use this model but idk how to make it work https://huggingface.co/Phr00t/WAN2.2-14B-Rapid-AllInOne
@RavirKun Hi - unfortunately, for Wan video it is not so easy just to "change" models. My workflow is completely optimized for 12 GB VRAM. It is a lot of work to find and test the right models and components to get fast and high quality outputs. Your linked model is a so called "all in one" model and seems to use a slightly different "architecture" with just one path. Best practise would be to start with the published workflow of this model and check if you get the basics running. Then you can try to optimize it, using components from my workflow for example: like blockswapping, multiplying, upscaling, Triton + SageAttention, options for different modes, etc.
@arkinson Ah, I see. Your workflow is working great — I just wanted to create some spicy content, and the default model/LoRA makes that hard to generate. So I did some research and found out I should use an "uncensored" model. But it's all good now since I found a LoRA that works.
I do have one issue I'm struggling to fix: How do you prevent the generation from looking like slow motion? All my outputs look like they're in slow motion. Someone suggested increasing the strength to 2–3, but when I use strength 3, my image gradually gets covered in red. Right now, I'm using 0.90 for both high and low in LoraLoaderModelOnly.
@RavirKun Just so we're on the same page - generally:
1. Please dont touch the settings for the default lightning Loras (lightx2v).
2. Additional Loras: The weight depent on your Lora Type Wan2.1 or Wan2.2. Please see my description in the workflow.
Slow motion: This can have various causes: T2V or I2V, prompt, start image, used Loras, etc.
I would start with some very simple tests like this: T2V, low resolution for quick testing, no additional Loras, simple prompt with "strong movement" like: "a man runs fast on the road", 5 second video. This should work out of the box.
Btw. I did some first quick tests with the nsfw model from your link. It looks interesting so far and I will give it some more tries.
@RavirKun slow motion effect is most often caused by NSFW Loras, especially those responsible for breast physics and other swaying motions. To avoid this effect, play with the Lora weight. It's individual, one need 0.8, other 2.0. Some Lora are basically incurable, just look for alternatives. Also, sometimes you need to describe the action as intense, fast, etc.
@hasenbein931 Thank you so much for your model review. I`m really impressed about the quality you get out with just 8 GB VRAM 👍
Stupid question: Can anyone give me a hint where I can find all the model reviews here on the model page??? I only get the latest reviews in the Notifications tab. Maybe I`m completely blind 🙄
[Edit: Ok - finally I found it 🙄 In the "Details" menue on the right just click on "Very Positive" right from "Reviews". This opens a list of all reviews]
@Bakazaya Thank you so much for buzzing and your feedback 🙂🙂
@arkinson Hi and thank you for that worflow. I'm a very first time user and still didn't have found a worflow that works with my setup. But when i try to launch it, some custom nodes are missing. I could find the JW ones on Git, but still haven't found the "FinalFrameSelector" nor the "Load Image Based on Number"
Thanks for helping
Have you tried - Manager > Install Missing Custom Nodes > Install All ?
@Louison73 Hi - as Bakazaya allready mentioned use the comfyui manager to install all custom nodes. Depending on your comfyui version you might have to install the manager manually first. If the manage did not list all needed custom nodes - just ask here again.
Thank you 2 for your answers. I check that now. I'm using Stability Matrix.
Done : no "manager" in this Comfyui version found. But it's the only one i managed to install and run
@Louison73 This is normal. You have to install it by yourself. Just google for comfyui manager. There are lots of easy to understand desriptions how to install and to use it.
Btw. I tested Stabillity Matrix for myself. The "idea" of an installation manager is not bad, but in the real world this software is extreme buggy and unstable and if you run into errors you get no help and you have no chances to fix it yourself. If you are a on windows I would recommend you "comfyui desktop for windows" instead. This version works very well and is regulary maintained.
@arkinson You're right : easy to install, easy to use, but bugged. Do you recommend this version ?
https://docs.comfy.org/installation/desktop/windows
@Louison73 Yes, that`s the right link. Installation on Windows is very easy, just execute the windows installer. You do not have to struggle with Python, Git, etc.
After installation I would recommend you the following steps:
1. As a beginner I would use the standard setting with automatic updates.
2. Install the comfyui manager for easy custom nodes installation. As a beginner do not try to clone from github, 99% works via the manager.
3. You might also have a look at the chapter "Adding extra model path" in the installation guide, if you like to organize all your models in a directory structure of your own, like a separate drive/directory for all your models.
Usage and maintenance is very easy and all works like a "normal" Windows program. Just a short hint: Settings and the very usefull Help menue are "hidden" under the small icon in the top left corner of the comfyui window.
@arkinson I did install it as you told me. The problem is that i do not want it to run on C: because of a lack of diskspace. Then i tried the portable vesrion on my other SSD, but it does not contain the manager.
If i use the desktop version, can i put the loras and checkpoints in another HD ?
Edit : Forget this, i managed to install it on another disk XD
Re-Edit : I encounter another issue. For example, I uploaded models in "D:\ComfyUI\resources\ComfyUI\models\unet". And loras and so. But whane launching ComfyUI Desktop, it doesn't find em. Pffff....
@Louison73 1. You could use the portable version too. It has pros and cons. Look at my small "tutorial" here at the end of the model page. Installing the manager takes 1 - 2 minutes. As I allready mentioned - use google.
2. You really should read and understand my last post, point 3 😉You have to configure the extra model pass yaml file. This is very easy. And just for fun: once again - use google 🙂
@arkinson i finaly managed to modify the extra model path ! now, i try my first generation
@Louison73 Good luck and happy generating🙂 and just to share your experiance for other users:
you use the windows desktop version now?
you installed the comfyui manager?
@arkinson Yes Desktop with manager. I made a first kitty video for testing. Now i'm trying i2v. But encoutered an issue : "WanImageToVideo.execute() missing 2 required positional arguments: 'width' and 'height'". But couldn't find the parameters.
Once again : thank you for helping !
@Louison73 Thank you for your feedback 👍 Seems the workflow found no start image. I would guess you simply havn`t selected YOUR local start image. Just drag-n-drop your local image to the image loader node. Make sure you have selected the right options for I2V and allways one option in the black switch nodes.
@arkinson It works ! But it seems that i have a lack of memory. After a generation of a 5 s video, i cannot launch another one without a "Reconnection" error
@Louison73 I have the same thing, just restart ComfyUi after each generation.
@Louison73 What means "it works" ???
@Bakazaya Any error messages? Latest comfyui version? All nodes up to date? Your swap file is large enough???
@arkinson
Yes, everything is updated, the swap file has been expanded. If I launch the next generation without restart, it crashes at the RIFE VFI point. After the crash, a red "Reconnection" window appears in the upper right corner, with no error descriptions. Restarting ComfyUi takes 15 seconds, it is not a problem to restart it after each generation. So I don't care. Otherwise, everything works fine.
@Bakazaya Hi - thank you for your feedback. I`m pretty shure it is just a RAM problem. Please could you run a test with a swap file size of 32 GB at least. Other way would be to reduce the "clear_cach_after_n_frames" to 4 for example. Just expand the RIFE VFI node to find the option.
@arkinson I mean that i finally managed to use it and generate 5s videos in i2v. But i think that may 16Go of RAM are not enough. Otherwise, i don't know what is the "swap file". Thanks again for this great job.
@arkinson Increasing it to 32 GB didn't change anything. It feels like the first generation takes up memory and can't be freed up during the next generation. Perhaps this is a personal issue with my system. I doubt it's a Workflow issue. I even reinstalled ComfyUi, and it lost the Manager after updating Custom Nodes, lol. Now that button is simply gone. I hate ComfyUi >_<
@Bakazaya i ordered 32Go RAM online. I'll see if it changes something and tell you
@Louison73 You had the same problems with 16GB RAM as I did with 32GB RAM. Something tells me 48GB won't make much of a difference. But in any case, those extra 32GB will be useful if you're doing AI generations. Even Stable Diffusion sometimes requires more than 16GB.
@Bakazaya I found some older issues with the RIFE VFI node and heavy RAM consumption, for example: Is this FiLM VFI running out of memory? · Issue #39 · Fannovel16/ComfyUI-Frame-Interpolation
Ok, I checked my system. The actual size of my swap file is about 56 GB 🙄 I use automatic swap file size (determined by the system). Set your swap file to automatic too. Keep in mind to use a drive with sufficiant free disk space and run a new test please.
@Bakazaya You might also test this:
1. Install custom node: "comfyui-easy-use".
2. Add node "Clear Cache All".
3. Connect the node`s input with the first or second "Video Combine" node output.
I´m not sure if this will help, but it could be a workaround.
Btw. My swap file size is over 100 GB after some generations 🙄🙄
@arkinson Thanks for the solution. 👍 The issue was with swap file mode. After changing it to automatic, the crashes stopped. I ran 10 generations in a row and everything was stable. So it looks like swap file mode is more important than size.
@Bakazaya I`m really glad that we found a solution 🙂I never thought this node could use so much RAM. The few users who have been affected by crashes so far were able to resolve the issue with swap files of approximately 24–32 GB. Thanks again for your feedback.
Is it possible tu use other checkpoints in this workflow ? The one you use are unet, beu what if i wanna use this, for example ?
https://civitai.com/models/1995784/smooth-mix-wan-22-i2vt2v-14b?modelVersionId=2260110
@Louison73 This is a "full" model, to run it you need at least an RTX 3090 with 24GB of VRAM, and better yet, a 5090 with 32GB.
@Bakazaya Thanks for reply. I keep my 3060 by now, but i will increase my RAM to 48Go soon. And see what goes better.
@Jio_R Video Combine Settings - I start a new comment here, just to keep track of the discussion. For the case, someone else is interessted in, please have a look at the original image comments here: https://civitai.com/images/110455042).
I did some more serious side-by-side tests with the following settings: pix_fmt: yuv420p vs yuv420p10le and crf: 19 vs 14 with my standard start resolution 720 x 480. To be honest, the visible improvement in quality is marginal (to nearly not visible), even in full sreen mode. Technically the file size and the bitrate is higher (up to 80-90%). I suspect that your improvements mainly result in your much higher starting resolution.
What do you mean with: "Setting Value FPS 24 Format video/mp4 or video/webm"??? and "Input color depth 16bit"?? 🙄
And what do you mean with: "...someone managed to solve the quality issue". I had a short look in one of the workflows of your link: https://markdkberry.com/workflows/research/#upscaling. As I can see, he is using VACE. Unfortunately I don`t know nothing about it and I did not read the whole article. So my question is, what is the advantage of this workflow in your opinion?
Yes, the original image was (1536x2048) https://civitai.com/images/110455042 I also increased the "Image Resize by Longer Side" in your workflow to 1024, used "lcm" + "beta" for sampler name + scheduler, left the "Video Combine preview video 16 fps" block unchanged, and changed the "Video Combine final video 60 fps" block to only 14 crf and yuv420p10le for pix_fmt. Regarding (setting the FPS value), this was a response from the GPT chat, so I simply quoted what he suggested. As for the phrase "someone managed to solve the quality problem," I dared to assume that perhaps his knowledge of his workflow could provide a more detailed picture, but in reality, it turned out that this is different and not applicable to your scheme. And yes, I noticed that with each new subsequent generation to continue the video from the last frame, the subsequent video becomes noticeably slower.
Oh my - I just wrote you a complete reply and accidentally clicked next to the window and whoops - everything is gone. I hate civitai 😡😡
Ok, I got it, the somewhat strange values where just Chat GPT outputs. I first thought you have some magic inputs 🙄
Sampler and Sheduler: Yes, trying other combinations can lead to "better" results. I did a range of test a while ago, but short conclusion is: Every combination will generate different outputs (under exactly same conditions) and it is hard or impossible to tell wich is best or better in general.
VACE workflow from your link: Yes it seems to be a mostly different "architecture" with other nodes and "special" VACE models, so I would say: no change just to "implement" it. My question was, if you have any practical experiances with VACE or some knowlage about it, cause it is a lot of work seriously testing a new workflow. I`m not sure if VACE will deliver better quality/speed/movements/prompt adherance etc. but it might be interesting for more advanced techiques like controlnet/masking etc.
@arkinson No, I'm just a regular user, just above a beginner level, just understanding. Unfortunately, I don't have as much knowledge as you. And of course, I simply outlined the process, in no way asking you to take it on. :) Your workflow suits me perfectly, considering how much I've previously struggled to get things running on my RTX 2060 GB graphics card. So thank you for this well-oiled workflow!
@Jio_R Yes, I understand you. And seriously, I learnd a lot for myself of users like you 🙂 I`m just very interested in trying new techniques or improving allready existing ones, like this workflow. The basis is, of course, to distinguish as clearly as possible between coincidences and reproducible results. My own knowledge and even my testing capacities for video generation is limited - that's why I sometimes ask very intense questions, before starting many try and error runs 😉🙂
VACE: seems to be also powerfull for lip-sync and different kinds of video to video generation...
@arkinson And yet, your hard work and dedication to achieving a workflow for low-power graphics cards brings previously generated images to life, which looks very cool. The VACE technology certainly has potential, but it requires knowledge of the architecture, and I'm not as proficient as I'd like to be.
@paulo Hi - thank you so much for all your buzzing 😋😋🙂
i have 3060 12 gb but only 24 gb system ram. does it work?
@prasanth_dma808 Hi - the workflow needs more than 32 GB RAM. A larger swap file size is always necessary. If you increase the size of your swap file accordingly, it should work. Please have a look here too: https://civitai.com/models/1852904/wan-22-workflow-optimized-for-rtx-3060-12-gb-vram-gpu?dialog=commentThread&commentId=992196
Great workflow ☺
This applies to all generative AI tools: it's best to have at least 64GB of main memory.
Excessive use of swap files shortens SSD lifespan (TBW) and slows generation speed.
※ Drive writes are significantly slower than memory access. Main memory durability is quite high and unlikely to fail unless shorted.
Thank you 🙂 In theory, you are right, of course, but in practice this is more of a minor problem I would say. And if the fairy godmother came from the fairy tale and I had three wishes, I would wish: 1. more VRAM, 2. more VRAM and last but not least: much more VRAM 😉😅🙂
I'm trying to use this workflow (2.2) but with a fresh install of comfy the workflow is coming all scrambled up.
I've chosen :
"select mode" text to video->no
"select mode" Image to video -> yes (still shows as muted )
"select mode" Single image -> yes (still shows as muted)
What am I missing to do I2V?
Hi - your selection is right for I2V generation. Just to be sure you look at the right subgroup/nodes, please make two simple tests:
1. toggle the "select mode" Image to video yes/no-> in the subgroup "01 Image to Video" the nodes should be toggle on/off (means not bypassed/bypassed).
2 . toggle the "select mode" Single image yes/no-> in the subgroup "01 Single Image" the node "Input First Frame Image" should toggle on/off.
If that doesn't work, please check the custom node pack "rgthree-comfyui" via the manager. This should be installed correctly and should not display any conflicts.
@arkinson Turning the "01 Text to video" group "select mode" node "text to video" yes/no, does enable/disable the entire 01 group.
Regarding I2V I only have a 02 group for that, the 01 group seems to be only for text to video.
For the "select mode" "Single image" inside the "02 Image to Video" group when turning it on, it only enables the node "Input first frame image". The entire "02 image to video" is still disabled/muted except for that one node. Plus the "select mode single image node" also remains muted.
rgthree-comfyui is not showing any errors on the manager. I'm on the 1.0.251127084 nightly version of that node pack
@requiemss The switch nodes don`t toggle the groups itself, they just toggle the nodes inside the specific group (main-groups or sub-groups). I guess this is your missunderstanding.
You didn't answer for test 1, but test 2 works as designed and switching between T2V and I2V also works.
Just drag-n-drop a start image, create a prompt and start a generation. This should work out of the box.
[edit: "Regarding I2V I only have a 02 group for that, the 01 group seems to be only for text to video". You have to distinguish between the main groups and the subgroups:
main group = "02 Image to Video". Inside this main grou you find the subgroup "01 Image to Video" for example.]
@arkinson Hi, and thank you for the replies so far! I'm going to start once more to see if I can find the issue. I just freshly imported the 2.2 workflow.
1- On" 00 Input" on Select Mode Enable 01 Text to video I'm turning it to "No".
2- On "02 Image to Video" I'm turning on both Enable 01 Image to Video and Enable 01 Single Image.
3- I'm dropping a 480x740 image on "Input First Frame Image".
4- On CLIP Text Enconde (Positive Prompt) I'm writing my prompt.
If I hit run I get "WanImageToVideo.execute() missing 2 required positional arguments 'width' and 'height'. The node that failed is WanImageToVideo.
I hope this helps highlighting what I'm doing wrong.
@requiemss Thank you for hanging on👍 I just did a fresh comfyui install on one of my systems too and it seems there is an issue with a custom nodepack with latest comfyui versions. Please check "Various Comfyui Nodes by Type" in the manager. I actually get a red alert: "import failed". This would cause your error too.
@arkinson Of course! I actually managed to sort that issue out. I was lacking some python psound or something library (this was before I even made the original post).
So sadly the issue is still ongoing for me and that node pack is fully installed, no red "import failed" error
@requiemss You are right - and for any other users with the not installable "Various Comfyui Nodes by Type" use this workaround: stop working with update · Issue #20 · jamesWalker55/comfyui-various.
OK, back to your error "WanImageToVideo.execute() missing 2 required positional arguments 'width' and 'height'. It simply means, that the node gets no image data form the "Input First Frame Image" node.
Make sure to select an image form YOUR local drive: use the "choose file to upload" option and select your file. Your image has to be displayed in the node immediately.
Btw. to drag-n-drop images into the node seems not to work after latest comfyui updates anymore 🙄 These updates drives me crazy 😣
@arkinson Hi! Yes, I'm providing an image into the "Input First Frame Image" from my local storage, through browse (not drag and drop) and it shows on the node display. But I still face that WanImageToVideo.execute() missing 2 required positional arguments 'width' and 'height' error :/
I just tried connecting the "Input first frame image" image exit directly to the "WanImageToVideo" node to see if maybe the switch was causing issues, but I'm still getting the same failure on that node. Super weird. I also tried a different image just in case.
Another quick update, by unhooking the getimagesize node from wanimage2video and providing width and height manually, the workflow is running. Atm I can't try the results, and I doubt I'll get something good since I'm not fully aware of if I should provide my source image values or something else. But the issue is there somewhere. Either that node is not getting the values from the image, not passing them or the image is being lost in the switch before.
@requiemss Thank you for your feedback. Thats really strange. Since the last updates, I've also been a little unsure whether everything is still working properly and suspected that currently only aspect ratios divisible by 16 are working. But on my system it doesn't matter, all aspect ratios of the start image are working.
What I can offer you is a relatively simple workflow just for I2V for quick testing. But I probably won't be able to send you the link until tomorrow.
On the other hand, please check again in the manager to see if there are any other conflicts with the custom nodes.
T2V is working??
@arkinson Happy to try any i2v workflow you send my way. I'll also try t2v tomorrow and let you know, it's a bit late here atm :)
@requiemss Hi - I opened a new article for easy testing, file exchange and further discussion. Please look here and try the simple test: https://civitai.com/articles/23042
I am getting error for below nodes in fresh ComfyUI_windows_portable_nvidia installation
JWIntegerDiv, JWFloatToInteger, JWImageResize
Anyone faced same ?
It works fine in Desktop version, issue is with portable version only
@baburaolele80 Yes - unfortunately it`s a known issue with this custom nodes and latest comfyui updates. For a workaround look here: https://github.com/jamesWalker55/comfyui-various/issues/20 or have a look here (just scroll down to the bottom for a short installation description for the embedded version: https://civitai.com/articles/23042/test-wan-workflow
Great worklow thanks :) !
@Jubien Thank you - and happy generating 🙂
Hi @arkinson! It's me again :). I've got a quick question, I've got a slightly more powerful video card (3080ti). Is there any optimization/change you would recommend? I'm trying to take this as a learning opportunity :)
You can test higher resolutions for example.
@arkinson Going to jump in on this comment as I was going to ask something similar: I'm going to be upgrading to a 5060ti soon, will this workflow just (hopefully) run faster on it or should I find one optimized for that card? I don't want to leave processing power/vram on the table
@dragon509127 Hi - I`m really not an "expert" for "high power" GPUs - so you better do not ask an old school Austin Mini driver what you can do with your brand new Ferrari 😂🙂 But generally I would say:
1. increasing resolution would give you mutch better output quality,
2. bypass the lighning Lora and increase steps - you might got better movements,
3 you might try better/larger models (fp16).
Or simply look for a workflow that is allready optimized for your gpu.
This is amazing! Thanks!
I was using Wan 2.1 before this and it took longer to do 3 seconds I2V in that than 5 seconds in this Wan 2.2 workflow! I have no idea how you figured all this out, but you're a wizard and I appreciate it. :)
Specs:
Ryzen R5 3600
32GB RAM 3200
3060 12 GB
@Flowares Hi - thank you so much for your feedback and I´m really glad this little automation is usefull for others too 😋🙂
Might look like magic but finally it is really simple: low framerate, low resolution and lighning Lora for fast generation and final framerate multiplying and upscaling for good quality and last but not least blockswapping for low VRAM consumtion 😉
The "tricky" part was more to find and understand these components, put them together, test and optimize the settings and of course, just to find the right models. Yeah - a lot of try and error.... 🙄
Last step was just to simplify the inputs with sliders and switch nodes for the different options and to create a simple und usable "flow-design", some beta testing to kill the last "bugs" and writing a "hopfully understandable documentation" 🙂
@arkinson Aah, thanks for the detailed explanation. It's all kind of a maze to me at the moment, with all those nodes and models that are around these days. I'm really impressed that you kept RAM and VRAM usage in the boundaries so that it doesn't eat up pagefile resources and keeps it all at a usable speed.
With the wan 2.1 I was using it basically ate up all my RAM, VRAM and also a lot of my pagefile resources. It also put drives at 100% at certain times., while with your workflow I don't think I have even seen it touch the pagefiles at all!
The sliders and switches are also a really nice touch, causing it to be more beginner friendly. This helped a lot in my case as a beginner and means I don't have to go searching for certain nodes that help workflows, because there are just so many options.
I'm using 512x512 images with 6 steps and 6 seconds long and it takes around 6-7 minutes, which I'm totally content with!
So, thanks again! :)
@Flowares As I started with comfyui about 2-3 years ago everything was easy - just a handfull of nodes and a simple environment - and of course, no complex video generation 🙂 But I am afraid that is the way the world goes round. The pro is: software developes rapidly, but the cons are: for beginners it becomes harder and harder to get in ☹️
If you are interested to understand the basics of the video workflow you might have a look here: https://civitai.com/articles/23042/test-wan-workflow Just download the wan22 I2V testflow from the attachments. It is my simple streamline workflow without all the more complex switch options and much better understandable and perfect for testing.
@arkinson Yeah, it's a good thing it all evolves so quickly so people can figure out optimal workflows. I'll just keep an eye on this page and see if you figure out more ways to optimize your workflow and just use your way, because it has already helped in a major way.
Thanks for the link as well, I'll have a look if I have time to dig deeper into how it all exactly works and increase my knowledge. :)
The workflow works great for text to video, but there are no switches in my mode select box. I'd like to use image to video. Do you know what would cause this? I've tried reimporting the workflow to no avail.
@sasquatcheggs Hi - make sure your rgthree custom nodes are correctly installed (no node conflicts in the manager). In the black rgthree swich nodes you can select yes/no. Just make your decision.
@arkinson Thanks for your response. I decided to try switching to portable Comfy. I grabbed a fresh copy and installed manager. I then installed all missing custom nodes and one failed because it couldn't import "soundfile". I installed that manually. Once all custom nodes were installed and working, I still had a blank "Select Mode" box. The prompt select and Triton+Sage boxes are fine. I'm not very good with Comfy yet, but that isn't your problem. I've had issues with custom nodes every time I've tried to use ComfyUI so it's definitely user error. Thanks anyway.
@sasquatcheggs Hi - thank you so much for your feedback. Actually I have a lot of trouble with the latest comfui updates for myself (conflicting/not working custom nodes, high vram usage/OOM errors with workflows wich works properly before, drag and drop /copy and paste do not work anymore, etc....) ☹️
The missing "soundfile" is a known issue since some days. Please have a look here (scroll down to the end): https://civitai.com/articles/23042/test-wan-workflow User @requiemss published a short instruction how he installed the portable version and the workaround - and it seems to work. And you might also have a look at the comments here with @sxcbutt181 - could be the same problem.
@sasquatcheggs Important: Turn off the new Nodes 2.0 if you had it enabled. The rgthree nodes don`t support it actually.
Having issues with both versions. With 2.2 I receive the following error: Model SamplingSD3 'NoneType' object has no attribute 'clone.' It seems a fresh install of ComfyUI portable fixed this issue for other users but it didn't for me.
I tried going down to 2.1 and keep getting this error: LoraLoaderModelOnly Model in folder 'loras' with file name '91 Lighning\Wan21_I2V_14B_lightx2v_cfg_step_distill_lora_rank64.safetensors' not found. I double checked if I downloaded the right files and their save locations.
These errors seem to pop up right after the WanImageToVideo step.
Hi - please use only my version v2.2. The "NoneType" error simply means the node gets no input. Most common error is - you simply not selected the models from YOUR local path. As desribed, click in every blue model loader node and select YOUR local model. As you described for testing my version 2.1 you definately did not do that.
@arkinson Hello. On my next attempt at v2.2 I fresh installed ComfyUI portable again and went into every blue model loader node on the workflow and clicked on them, making them say either none or undefined. I then redownloaded the required downloads and saved them in their proper locations, restarted ComfyUI and went back into each blue model loader node and selected my local models as shown in your jpeg (although the Lora Loaders don't have "90 video\" before Wan21_I2V_14B_lightx2v_cfg_step_distill_lora_rank64.safetensors. Does that matter?)
Only Enable Image is active and Sage Attention is disabled as I have not installed it. I still get Model SamplingSD3 'NoneType' object has no attribute 'clone.' What am I doing wrong?
Text 2 Video works without issue
@sxcbutt181 Hi - regarding to your questions I see a lot of mismatch and confusion. So let`s go step by step.
1. If you see something like "91 Lighning\Wan......" in the model loader nodes you still try to use MY path and you did not choose YOUR local model. As discribed, you allways have to use YOUR downloaded models - even for additional Loras. If you can`t select YOUR downloaded model you might saved it in a wrong path for example.
2. T2V works - it means you have selected the right models for this part 👍
3. I2V: Once again, check your models. Check your first frame image - you have to select YOUR local image. Check you have selected a prompt field with a prompt in it of course. Deselect all additional Loras at first.
And please have a look here: https://civitai.com/articles/23042/test-wan-workflow. With the latest comfyui updates some custom nodes seems to have issues. At the end of the article you will find a short workaround for the portable version.
@arkinson Hello again
1. I don't see "91Lighning\" anymore after selecting my local model as you instructed. I only see "Wan21_I2V_14B_lightx2v_cfg_step_distill_lora_rank64.safetensors" in the High and Low Pass Lora Loaders. I even removed the space where the bouncing Lora was.
2. Yes, the process moves through ModelSamplingSD3 immediately in t2v. The issue must be in the 03 Image 2 Video Group right?
3. I don't get this error in your wan22 I2V testflow with these i2v models so maybe they're not the issue? I've dragged and dropped and clicked "choose file to upload" to put local images into the Load Image Node. They even show up in my ComfyUI input folder after. Yes, I use the same prompt field as t2v but with a different prompt when attempting i2v.
If you are referring to the 9 step workaround I already do that when fresh installing ComfyUI because I would also be missing the pysoundfile that causes failed JW node installs. Regarding your wan22 I2V testflow, while the process does go all the way through the workflow without this error, there are no videos at the bottom of the Preview Video and Final Video nodes. Also, the mp4 files that come out of it do not have a thumbnail in my ComfyUI output folder like the t2v ones do and I can only play them in VLC Media Player. Would this have anything to do with my error in the main workflow?
@sxcbutt181 Hi - thank you for testing and your detailed description👍 The difference between the test workflow and my published workflow are only the switch options - wich using the JW nodes.
"3. I don't get this error in your wan22 I2V testflow". Ok - from my point of view it seems you have a similar problem like @requiemss. (He solved it with a fresh install finally). Something must be completely wrong in your system, probably corresponding to the JW nodes. That could possibly explain your other problems as well. However, it is difficult to judge this from a distance and if you are not able to solve this by your own research I would say the more easy way for you might be a clean new install of your comfyui system. You might use requiemss instructions for a manual installation or have a look here: https://civitai.com/models/1852904/wan-22-workflow-optimized-for-rtx-3060-12-gb-vram-gpu?dialog=commentThread&commentId=1031214. @Icey64 link for the "One Click" installer looks interesting and seems to work properly. Even the installation of Triton and SageAttention seems to work.
@sxcbutt181 Turn off the new Nodes 2.0 if you had it enabled.
@arkinson I'm a damn fool. I was using v1.2 thinking it was 2.2 this entire time! The downloads and workflow itself work perfectly on Easy Install ComfyUI. Thank you for being so helpful and for the workflow. This is amazing
@sxcbutt181 Hi - thank you so much for your reply. Oh my, mismatching releases - yeah, we all get older day by day 😂 I´m glad you got it running 🙂 And good to know, that Easy Install ComfyUI seems to work out of the box. Have you installed Triton and SageAttention too??
@arkinson Yes, Triton is included and SageAttention was a quick and easy install. Improved timing by 25%
@nov1ce Thank you for buzzing 🙂
What?? I've never had a workflow for wan22 work perfectly from the first try! TYVM!!
@Icey64 Hi - thank you for your feedback and thank you so much for buzzing too 😋🙂🙂 I`m really glad it works well for you.
Please, could you tell me wich comfyui version and update version you use, cause a few new users seems to have heavy problems to get it running. Maybe comfyui beginner mistakes, but it seems there are some cruel custom node conflicts after the latest comfyui updates (later then comfyui version v0.3.68).
@arkinson I'm running comfyi 0.3.75, I installed the comfyui-easy-install which include sage attention and torch but I don't even need to use them with your wonderful workflow!
@Icey64 Thank you for the link 👍 That`s very interesting. So that seems to be the portable version. Did you a fresh install with latest version v0.3.75 or did you an update from an older version? Could you install all necessary custom nodes via the manager without any issues? Cause some users including myself had to run a workaround with missing "soundfile" to get all custom nodes working after the updates.
[edit: And I2V works too on your side? Sorry - lot`s of questions 🙄 but I try to catch the latest issues.]
@arkinson yep, it was pretty much a fresh install. After trying many workflows with sage and failing I found yours and as I said it worked straight away! Just installed missing nodes through manager and models you linked and was off to the races trying i2v straight away!
@Icey64 Thank you so much for confirming this 👍
@arkinson hello, I sent you a DM with a question about the multiple prompt boxes, I was just wondering what they're for? Thanks!
@Icey64 Sorry - I don`t use DM. The multiple prompt fields are just for conviniance and quick prompt development: Generate with prompt 1, while creating prompt 2 for example, or just use it as a place to save different prompts and quickly switch between them, etc.
Hello,
I keep having this issue. Any suggestions?
[Edit -Solution: Turn off Nodes 2.0 in comfyui]
What kind of issue? Your link did not work.
@arkinson sorry about that. The Select Mode node seems bugged. It looks like every option is stacked on top of each other and I can only select/deselect T2V.
@pitsnreact1 Seems an issue with custom nodes (JW nodes) after latest comfyui updates. Have a look at the last comments here and at this article: https://civitai.com/articles/23042/test-wan-workflow. A few user reported similar problems.
I really would like to see your screenshot, cause I have no glue how this looks like. Can you please write/open an article? Just give it a name like test, attach your image as a zip file and announce me in a short message.
@arkinson done!
Attention: Please turn off Nodes 2.0 in comfyui!
Some custom nodes used in my workflow actually don`t work with the new Node 2.0 setting. If you enabled it, please switch back to the old settings.
btw, this workflow works well with Q8 models on systems with 12GB of VRAM and 32GB of RAM. The only thing I changed from the default settings was bypassing block swap. As a result, a 6 sec video on the Q8 at 720p takes 7-9 minutes to generate. The difference compared to the Q4 is about 1 minute.
@Bakazaya Hi - do you talk about T2V or I2V, 6 steps and 720 x 480??? A generation time of 7-9 minutes for a 6s video is extreme fast. Wich GPU did you use?
Wich Q8 model did you use? Did you compare the Q8 model directly with the Q4 model? Did you notice any improvement in quality???
I am asking these questions because I also ran a few short tests some times ago. However, I could not detect any significant differences and assume that a resolution of 720 is far too low for the larger models
@arkinson I only use I2V models, can't say anything about T2V. I usually generate 560x720 for 6 sec, which takes 10-11 min. I've also noticed that after 3-4 generations, the time increases, and after the 10th, it can take up to 14 min. I have rtx 4070ti 12gb.
In my video with catgirl and revolver, I marked a checkpoint. Checkpoint page has a link to huggingface, and I downloaded Q8 from that link. Compared to Q4, the difference is quite noticeable, video quality is slightly better, but animation quality is way better. There are no choppy animations that were common in Q4, and overall, it seems to understand the prompt better. Comparing time difference and quality difference, I'd probably prefer Q8. But that's in my case, where the difference is only 1-2 minutes.
@Bakazaya Thank you so much for your explanations 👍 That`s very interesting. If I remember right, I only tried the Q6/Q8 "base" models with T2V. I found your model link in the video and will give it a try too. I am keen to see if my RTX 3060 can handle it.... Your quality looks great. But I can not compare it cause I never had tried animation for myself.
@Bakazaya Really - I don`t get it 🙄 I just did some quick tests with your linked Wan22 I2V Q8 A14B models (https://civitai.com/models/1820829?modelVersionId=2060527) and the first results in comparison to the Q4 models are:
- visibly better quality - even for realistic videos,
- seems to use LESS vram 🙄🙄,
- much faster generation time: about 9 - 10 minutes instead of 15 minutes for a 5 second video 🙄🙄🙄
All tests done with my RTX 3060 and my standard test settings (see my workflow description). I just finished a 10 second video without any issues. I think I'll better go to bed and try again in daylight and without any drugs 🙄😉🙂
@arkinson Also, I think I've found a way to improve quality a bit more without losing anything. I tested a bunch of different Lighting Loras, and some of them improved the video quality, but animations suffered. Either the motion was too slow, or they didn't understand prompts well.
Wan21_I2V_14B_lightx2v_cfg_step_distill_lora_rank64, the one you suggested in workflow, seemed like the best option, even though it sometimes degraded the quality. So I decided to look for an equivalent and found lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16. Animation result and prompt understanding are on the same level with rank64 lora, but with fewer distortions and artifacts. Link: https://huggingface.co/Kijai/WanVideo_comfy/blob/main/Lightx2v/lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors
@Bakazaya I just saw your comment. Very interesting, that you tested lightning Loras too. Even I tested a lot in the past I will check your link soon.
Speed tests with Q8 models: Yesterday I accidentally did all tests with Triton + SageAttention ON insted of OFF. So first I will repeat the simple comparison tests between Q4/Q8 to check speed and vram usage without Triton + SageAttention - cause that is the generall usecase for most users.
For quality tests between Q4/Q8 and lighning Loras I will create a simple I2V workflow for direct side-by-side view after the speed tests. As soon as I'm ready, I'll send you the link. It would be great if you could also run some tests.... 🙂
@Bakazaya I finished the speed tests (without Triton and SageAttention). You was right. Same thing on the RTX 3060: the Q8 model needs about 1 - 2 minutes longer for a 5 second video and finally there is nearly no difference in vram usage. So this looks really good for small sytems 🙂
I published my test-workflow for side-by-side comparison and started first quality tests. If you like, just download the "wan22 I2V q4 q8 lighning test.json" file from here: https://civitai.com/articles/23042/test-wan-workflow. Enter the models you like to compare in the first and second run-line.
@arkinson I downloaded this and tried, but at KSampler High, it throws the error 'no module named sageattention'. It looks like it requires triton and sage. Maybe I'll install them later.
@Bakazaya Yes, without SageAttention just bypass both nodes after the model loader nodes.
It would be really great if you can run some tests too, cause my own capacities are very limited 😕 I did some first "quality" tests yesterday and the results are quite sobering.
I compared "Q4 model + rank64 Lora" with "Q8 model + rank128 Lora". Lora weights: high = 3,0 and low = 1.5. In my few "photo realistic" test I can`t see any differences in quality, just slightly different movements.
Another way is to reduce the Lora weights to 1.0 and 1.0 for example. Now you get better prompt following and better/faster/more movements, but unfortunately the quality is significantly reduced, with more artifacts, blurred movements, etc.. 🙄 Maybe the golden shot is somewhere in the middle....
@arkinson Ok, I'll try bypassing that. Regarding quality, models are trained on "photorealism", so it's no surprise that it turns out well. The difference is more noticeable when generating anime or something unrealistic. For example, with rank64, unrealistic characters often have distorted eyes, like this one - https://civitai.com/images/112998821 . According to my tests with rank128 this happens much less often.
btw, I updated Comfy to latest version to fix Node Manager, and now for some reason my VRAM usage has dropped. On the old version, it used 9.5-9.8 GB during Sampler stages, now it's only 6.5-6.7 GB. It's hard to say how this affected the generation time. It felt more random, same thing could be done in 8 minutes or 12 minutes. I also saw numbers below 400 seconds for the first time. Once, a 6 sec video was done in 390 seconds, and I was like "wtf..." 😅
@Bakazaya You are on v0.3.76 now? It seems the latest updates have mismatched everything 🙄
Maybe I`m wrong, but in theory I would assume a larger model should output better quality or give better prompt following, etc. But I really guess, the limit is the use of the lighning Lora on the one hand and on the other hand the "low" resolution of max 720. A user with much better gpu used my workflow to generate with 1024 x 1024 and improved quality significantly....
@arkinson Yes, v0.3.76. With Triton and Sage blocks disabled, test workflow runs fine and I made several tests. The results confirmed my previous observations. Q8 and rank128 produce more consistent results than Q4 and rank64. Video quality itself is almost identical, but Q8 and rank128 produce fewer artifacts and distortions in complex scenes and follow complex prompts better. Regarding the time, Q4 is still only 1-2 minutes faster than Q8.
I'm a little disappointed that only half of VRAM is being used. This resource could be used to reduce rendering time or improve quality. But now I'm wondering, if only 6.5GB of VRAM is used, does that mean that an RTX4060 with 8GB can handle even Q8?
@Bakazaya Ok, in photorealistic scenarios I´m still not shure if I get better or worse results, but I will do some more tests.
Low vram usage: seems you use blockswap???
@arkinson No, I'm still bypassing block swap node.
after update comfyui, this nodes become like this. is this harmless? idk how to fix
My block swap looks the same, I just bypass it. Everything works fine without it, even feels a little faster.
@Bakazaya I got error if I bypass it. can u share ur workflow?
@RavirKun I'm using default v2.2 workflow, haven't changed or added anything. You're probably getting errors from Windows swap file. Try WIN + R, type sysdm.cpl, find swap file settings, choose the drive where you have ComfyUi installed, and set it to automatic mode. Save settings and restart your PC. (Ideally, there would be at least 30GB of free space on this disk)
@Bakazaya I followed your guide, but I still get errors if I bypass it. When I leave it alone (without bypassing), it works normally. But after generation, I have several issues — for example, the video doesn’t appear in the preview, and the generated video doesn’t play. Tried to revert comfyui version and no luck
@RavirKun Hi - have you turned on Node 2.0 setting in comfyui?? Please, switch back to the old node system via the comfyui menue.
@arkinson yes, I already turn off Node 2.0, everything is fine. only WanVideoBlockSwap nodes is broken.
@RavirKun Ok, seems you are on comfyui version 0.3.76? There is an open issue for the node: The Blockswap node interface is showing display issues. · Issue #9 · orssorbit/ComfyUI-wanBlockswap. (On my v0.3.75 it still works). The strange thing is: disabling blockswp should work without any other issues (exept OOM on vram with high loads like high resolution for example).
I will try to install a test system with v0.3.76 soon.
@arkinson all good now bro, it's fixed by deleting "nodes_nop.py" in comfy_extras folder.
@RavirKun Thank you for your feedback 👍You are still on version v0.3.76?
@arkinson yes, i'm still on v0.3.76 and everything was fine after I delete that file
JWFloatTolnteger
JWIntegerDiv
JWImageResizeByLongerSide These nodes are not installing, what is the link to download manually?
@paulo Hi - try the following: open Custom Nodes in manager and search for "jw". Install the "Various ComfyUI Nodes by Type". The github link is: jamesWalker55/comfyui-various
https://github.com/jamesWalker55/comfyui-various/issues/20 - python -m pip install soundfile
@paulo Yes, @rumbu711 is right. Use the workaround.
@rumbu711 Yes, that solved it
@rumbu711 This is the key, I was having problems and this fixed it
I followed all these steps and it keeps refusing to install
What am I doing wrong? the soundfile exist and is properly installed but these nodes will just not install for some reason
@Setian91 You might look here. This is allready linked in the FAQ. Please give a replay if it worked for you.
@arkinson I'm a little step further installed via CMD command in the comfyUI folder
But now I cant find the RIFEinterpolation node, ComfyUI manager can't find it and Google has poor results so far
@Setian91 Sorry, I don`t got it. What do you mean with: "installed via CMD command"? Do you used my link - or what we are talking about?
@arkinson Sorry manual install in Windows CMD commands withing the ComfyUI Custom Nodes folder
It was this package I couldn't find:
https://github.com/Fannovel16/ComfyUI-Frame-Interpolation
I found this is another workflow on CivitAI
I made it work, however I run into one issue, my video's are 10 seconds instead of 6, trying to figure that out, didn't change the settings btw
@Setian91 OK, you may do what you want. My question was if my link was helpful for you to solve your JWFloatTolnteger issue, cause this was your origanal question and this is the topic here.
@arkinson Sadly no, installing manually via a CMD command directly in the Nodes folder with git clone was the only thing working for me
@Setian91 You are completely on the wrong topic here 🙄
太麻烦了，对我们小白来说太不友好了，不知道怎么下载，管理器没有我不知道怎么能了，
"Warning: Ran out of memory when regular VAE decoding, retrying with tiled VAE decoding."
After trying First and Last frame, I have a 5070 and 32gb of ram ddr5.
On Windows: set your swap file size to automatic on a drive with at least 100 GB free space.
Ok, I am stupid, it was as easy as increased my SWAP file, I set Minimum 32gb, Maximum 64gb and it worked perfectly
@arkinson Yes, I did that, that solved my problem, thanks
Works great, i had no major issues.
Hi - thank you for your feedback 🙂
i need help... whenever it starts upscaling, it silent crashes at the rife vfi node. i dont finde rife vfi anywhere to download. you guys have any idea? using linux by the way...
@heiopei It`s a know ram problem (the node itself works well). The workflow needs much more than 32 GB ram. On Windows it allways helps to set a much larger swap file (automatic size on a drive with at least 100 GB free space). This should work with linux too. Another option is to reduce the setting "clear_cache_after_n_frames" from 16 to 4 for example. You will find the "RIFE Video Frame Interpolation" node by moving the covering nodes away in the output group.
@arkinson thanks for the reply! started tinkering a little bit with my very little knowledge :D 3sec are working fine almost always. 4sec sometimes and i almost always crash with 5 sec. I will try your tip change the settings later! thanks a lot! getting new ram is not an option with the market right now :D i'm really glad your workflow is relatively easy to understand and use and gives me the option to start using wan!
@heiopei As mentioned, no need to buy new RAM (I asume you have 32 GB installed). Just increase your swap file to about 32 - 64 GB at least, or better to automatic mode. This should be no problem with linux too. And use my hint to reduce the "clear_cache_after_n_frames" setting.
@rockedt Hi - thank you for buzzing 😋
I got problem when loading this workflow, the Select mode and all similar node like this are empty. I already installed rgthree-comfy https://postimg.cc/Zvv64392
You turned off Node 2.0??
@arkinson I haven't turned it on yet. But the problem happened randomly because I use cloud compute, some instance got problem, I need to hire a new one eventhough they use the same template
So it would be better if you release seperate workflow for t2v and i2v that not use the rgthree node
@AI_Imaging Use latest comfyui v0.4.0 and make sure you have no node conflicts. It doesn't matter whether it runs locally or in the cloud.
For a simple I2V workflow you can use my "wan22 I2V testflow.json" from here: https://civitai.com/articles/23042/test-wan-workflow
I’ve got the same problem, did you find any solution? The problem is with rghthree node, the fast group muter only doesn’t work, the bypasser works fine.
@bobj84658399 problem come from swwan, just uninstall that custom node
@AI_Imaging What is swwan??? 🙄
@bobj84658399 Latest comfyui version, all custom nodes up to date??? Check for node conflicts.
@arkinson it is in comfy manager, I read somewhere else about that and just remove it can fix the missing toogle. If someone facing the same problem like me, just look for swwan node and uninstall it. I also can not find that node name if problem not happen
Here you also can see it is swwan, it is not the real node type
@AI_Imaging Thank you for your explanation👍 Arg - I had only googled for swwan, wich gives a lot of wrong links. Googling "comfyui swwan" directs to the "swwan" node.
Hi, this is a really great workflow — thank you for sharing it.
Right now, I’m only able to generate clips that are about 3 seconds long.
If I’d like to increase the duration to around 6 seconds, what should I adjust in the workflow?
I’ve tried looking for the relevant setting, but I can’t seem to find where the clip length is controlled.
Any guidance would be greatly appreciated. Thank you!
@Saxlive Hi - thank you.
Do you see the brown slider nodes in "00 Input" group??? If you can`t see/use the "Clip Length (in seconds)" Slider, check the manager for custom node conflicts. Make sure you have no issues with the "ComfyUI-mxToolkit". And of course, turn off Nodes 2 in comfyui.
@Squints Hi - thank you so much for your feedback and your buzzing 😋🙂
JWFloatToInteger ,JWIntegerDiv, JWImageResizeByLongerSide
I cant for the life of me fix these nodes. I tried updating with the manager, deleting it, also tried to manually delete the folder and reinstall through command prompt. They are still missing. any solution?
@MakingNasty123 Look at the comments first 😉 https://civitai.com/models/1852904/wan-22-workflow-optimized-for-rtx-3060-12-gb-vram-gpu?dialog=commentThread&commentId=1037467
Yes, take a look at the comments below. There are some solutions to the problem, and they work.
I also just tried to fix this situation using GPT. I sent the launch log, and as a result, it suggested: "You need to reinstall all node dependencies in the current Python that ComfyUI Desktop is using"
Default installation path:
C:\Users\username\Documents\ComfyUI\.venv\Scripts\python.exe
In CMD:
And it worked. Disable ComfyUI automatic updates :) lol
Comfy version:
Desktop 0.6.1
Python version: 3.12.9
In workflow: wan 2.2 i2v t2v FLF SA v2.2
Huge thanks to the workflow author. The best.
@WhisterB_ART Thank you for your feedback. I`m glad you got it running - and happy generating 🙂🙂
This workflow is the best. Just one problem is sometime when doing upscale and interpolation, comfyui just crashed with my 16gb nvram 5070 ti. How do i optimize the setting for low vram, it can take more time but dont crash the app. Tks
@AI_Imaging Look at the comments first 😉 We had often discussed this before. Increase your swap file. And you didn't answer my question here: https://civitai.com/models/1852904/wan-22-workflow-optimized-for-rtx-3060-12-gb-vram-gpu?dialog=commentThread&commentId=1040662 Comments and solutions should be usefull for others too.
I run it over the cloud and it is linux distribution. Does anybody know the proper way to increase the swap file in linux
@AI_Imaging You might google the Linux commands, but I`m not sure if this will be a simple task on a cloud server. Another option would be to rent a cloud server with at least 64 GB RAM or more. And keep in mind, these crashes are a RAM and not a VRAM problem.
And use my hint regarding "clear_cache_after_n_frames" - look here: https://civitai.com/models/1852904/wan-22-workflow-optimized-for-rtx-3060-12-gb-vram-gpu?dialog=commentThread&commentId=1039806
@AI_Imaging Thank you so much for buzzing 😋😋
It would be awesome if you can release an infinite loop video, i see it can extract the last frame and it could be the first frame for the next clip, in 1 task only
@AI_Imaging You can expand your videos using the lastframe image as start image for a new generation, but you have to merge your videos manually.
In the video combine/save nodes you can use the loop option - this automatically outputs a "looping" video. It doubles the video length, so the generated video runs first forewards then backwards. Works quite well for simple infinite loops.
Or you simply use the first- to lastframe option with the same image.
生成其实成功了，但是进度条卡在93，每次都要从新打开才能解决，主要卡在了RIFE VFI这个节点上，绕过了它生成的视频就是60帧的快速视频
clear cache after n frames 设置到了4，虚拟内存也增加了
@litianfan800797 Yes, as allready often mentioned here, these crashes are a RAM problem. Setting "clear cache after n frames" = 4 helps and increasing swap file helps too.
It works, but the quality is terrible, the characters' eyes are just blurry, I saw another workflow for images and it worked faster and better
If so, which one is it?
Workflow is awesome and I can generate 5 sec I2V vid in 8 minutes but the quality is bad especially the pixelation in the face region when there's lots of moving like walking. Any tips?
@nitrovtx Hi - Merry Christmas 🙂
T2V or I2V?? Do you use additional Loras? Could you provide a sample video with prompt, Loras, setting etc.?
You might test other lighning Loras too. We allready had some discussions here in the comments.
@arkinson Happy holidays!
Video posted by nitrovtx
Here's a sample video, it's I2V. You can see the distortion on the eyes and the swimsuit. Maybe not as bad as I made it out to be. The swimsuit can be fixed by removing the pattern, but the eyes are just too noticeably bad for me.
Aside from the bouncy walk loras everything else is on default settings, all the models I downloaded from the workflow.
@nitrovtx A video says more than a thousand words 🙂 Thank you for uploading it. I would say the distortion is not sooo bad as I assumed. Similar things I see in my videos often too and sometimes it is a lot of try and error to get one good result.
What I would try, in the following order, is:
- trying other/different start images and prompts, generating with different seeds,
- lowering the lora weights of the bouncy walk lora - I assume you use 1.0 and 1.0,
- lowering the weights of the lighning lora,
- testing the newer wan2.2 lightning loras - but believe me, that`s a lot of try and error work and if you find a solution for one case it is not sure it will work in another case too.
And I've got one more 😉 disable the lightning loras, generate with 20 steps or more and go and drink a coffee or a beer or both 😅
@arkinson Thanks for the replies! I've tried different images and adding details like red eyes, striped swimsuit, etc. but the problem still persists. I'm going to try all your other suggestions and share the results if I find the solution. That last one though 🤣
@nitrovtx Have you tried a different sampler? It's set to Euler by default but I've had better results with ddim. Also have you tried increasing the number of steps? The RIFE VFI also seems to exaggerate those artifacts since it's essentially extended the artifact in one frame to multiple frames; worth bypassing to see if the lower FPS is a decent trade off.
@NeilHamburger Umm... stupid question, but how do you bypass the RIFE VFI or anything else for that matter? This is actually my first time using ComfyUI. Edit: is it just by right-clicking and selecting bypass, and that’s it?
@nitrovtx Unfortunately I forgot one point that might be important for you. Recently a user reported simillar issues generating "anime style" videos. He switched to Q8 models and another lightnig lora and it seems he got better results. For myself I have no experiances with anime videos but you should have a look at this. (The Q8 models works well with 12 gb vram too). For more details look here: https://civitai.com/models/1852904/wan-22-workflow-optimized-for-rtx-3060-12-gb-vram-gpu?dialog=commentThread&commentId=1034251
And yes, trying @NeilHamburger hints also makes sense. To bypass a node just click "bypass".
@arkinson So I tried the Q8 models and the lightning lora you suggested, using the same image and prompt and I can immediately see improvement in quality.
The lighting on the skin and the animation is especially impressive! but the artifacts and distortions are still there though not as distracting, I might have to try the bypass suggestion from @NeilHamburger.
Thanks for all the help!
@nitrovtx Have you tried other Loras? I just noticed that the eyes with the bouncy walk Lora also appear a little blurry on my videos - even with a realistic start image and few motion: https://civitai.com/images/110173521. If you have enough VRAM, I would definitely try higher starting resolutions too. And prompting like "detailed/highly detailed face" might be helpfull.
@arkinson Yup, tried https://civitai.com/models/1363473?modelVersionId=2254373 and got similar results.
Also, I tried the RIFE VFI bypass but increase the fps to 24fps for smoother motion and what do you know, it solved the artifact issue AND the slow motion... While also introducing new problems...
For example, https://civitai.com/images/115244178 - At the 3-second mark, she starts walking in place, and the camera stops moving. It worked fine at 16fps... BUT if you extend it beyond 5 seconds, the same issue occurs (it happens in your video too).
While here, https://civitai.com/images/115245096 (NSFW Warning!) - Almost as if it’s playing at 2x speed or something. Again, no problem in 16fps but it's noticeably choppy especially in the bouncy walk one.
And so, I lowered the lightning LoRA weight to 1 and 1, and while it kinda fixed the 2x speed issue, the overall motion feels stiff, robotic, and unnatural. You can definitely tell that it’s a generated AI video.
I wonder if there is something wrong with my workflow setup. https://imgur.com/a/05sAvqm
this is what I change to make it 24fps
@nitrovtx Oh my - you did a lot of testing allready. Thank you so much for sharing your results. I really appreciate it 👍 cause my recources are limited too and that`s the way what improves the knowledge here.
Just a few words in advance: I wouldn't expect too much from Wan video generally, especially not from this heavily stripped-down 12 GB VRAM version here. In most cases it works pretty well, but sometimes it is more easy not to "force" the AI, jus skip your idea and try something completely different 😅🤣
But if you like experimenting and testing the limits - here we go 🙂
Please use my simple I2V workflow: https://civitai.com/articles/23042/test-wan-workflow download: "wan22 I2V testflow.json". It is exactly the same workflow but without the more complex switch options and in an easy to understand and to modify "design".
To narrow down the artifact and blurry eyes problem I would suggest some simple systematic tests: Use your worst case start image and prompt and run some short generations:
1. without any additional loras. If that doesn't help
2. disable the lightning loras too and run with at least 20 steps.
I would guess the first problem are additional loras, then the lighning loras and then start image/prompt. (You also might try some realistic start images). In my expreriance the combination of lighning lora and bouncy walk lora causes regulary some problems like: looping movements, forewards/backwards movements, artefacts, etc.
I can`t open your workflow link ("Imgur is temporarily over capacity. Please try again later."). To test higher frame rates: just delete my "clip length" slider and frame calculation and set it manually, for exampel: 3 [seconds] * 24 [fps] +1 = 73 [frames]. In the "Video Combine" preview node set 24 fps and in "Video Combine" output set 24 * 4 = 96 fps. Otherwise you get fast speed movies.
"Your" eyes here https://civitai.com/images/114911453 are really bad 😅 There are a lot of much better videos on this lora page, but most of them were generated with another model. I allready tested some of these "combined" models (lighning lora + other loras allready included). In my experiance the problem is: they are working very well for some special use cases, but if you try to use them with your own loras it often goes completely wrong and you never get what is allready under the hood.
Uhhh - that was more then two words. I hope you will find some inspiration 😂🙂
@nitrovtx I just found a mistake in my workflow, which may have contributed the issues (I didn`t splitted the generation steps over both KSamplers). And I found an interesting model which seems to solve the speed and quality issues. Yes it is a "combined" model, but first quick tests yielding astonishing results.
So if you like, please download my new testflow "wan22 I2V testflow neu.json" from here: https://civitai.com/articles/23042 and the new models from here: https://civitai.com/models/2053259?modelVersionId=2520805 I actually use the Q6 models, but Q8 should work too. Keep in mind not to activate the lighning lora in combination with these models.
@arkinson Sorry 😔, I haven't had time to do more testing, but it seems like there's more coming! I'll keep you posted when I do.
@arkinson Yeah so, I finally did some more testing. Using your test workflow and the models you suggested (I used Q8). Although not as bad, the artifacts and distortion still there but it also changes the character's face https://files.catbox.moe/a6iuzy.mp4
Just like my previous comment, it seems like frame interpolation is the culprit. So, here's some test with no frame interpolation but I again increased the fps to 24 using this model https://civitai.com/models/2190659?modelVersionId=2466604 it's a combined model also
https://files.catbox.moe/c1rk4p.mp4
https://files.catbox.moe/3xvyl5.mp4
If I make it longer than 3 seconds it breaks, and she will just walk in place. Lora limitation I guess, other loras works just fine regardless of fps and video length though I still have to do more testing, but I'm satisfied for now.
@nitrovtx Hi - thank you for hanging on here 🙂 Ok - increasing framerate might help. I would also try higher resolutions (set BlockSwap = 40) and see how far you can go. Did you some tests with the standard model and without any lighning lora already?? I would guess this could be important.
It seems you run all test with the same anime style character. Maybe it would help to test some completely other characters too. And if you stay strong to anime style I would have a look for special models and lightning loras - but as allready mentioned, I have no experiances here.
If you like, try my next "wan22 I2V test 2025 12 29.json" workflow: https://civitai.com/articles/23042 I used the "PainterI2V" node here. It seems to give better movements. I will do some more tests over the next days and if it works well I will publish a new "official" release of my workflow.
I had a short talk with the Q6/Q8 model creator yesterday: https://civitai.com/models/2053259?modelVersionId=2520805&dialog=commentThread&commentId=1058864 His hint with the black start image is really cool.
You definitely should try his multi-scene workflow (for the link scroll down to the bottom of the model description). This solution is brilliant. You can create longer very congruent clips by multi short clips (each 5 seconds for example) in a semi-automatic process. This definitely helps against "looping" movements you allways get with longer single clips. Next I will dive into SVI V2 Pro. I just had a short look, but it seems to be very interesting.... Oh my, 8 weaks ago I was lucky to get Wan22 running in general, but now everything runs much to fast here 🙄🙂
@arkinson Happy belated New Year! Sorry for the radio silence—this month has been hectic and will continue to be, so I don’t have time for more testing right now. I’ll share my findings here whenever I can. Again, my apologies.
@nitrovtx Yeah, this year is almost over again 🙄😅 I was completely out of brain for some days catching a heavy cold 🤧. I wish you a Happy New Year too. Don`t hurry. Your tests so far were reason enough for me to take a critical look at my previous workflow. I just published a new version v3.0. It contains mainly the things we allready discussed in the test workflows - so it will not really solve your issue. But I believe it will significantly improve the work for "common" users. If you find some news on your way, please give me a hint and I will have a look on it 🙂
Take note: If you have the issue with "JWFloatToInteger ,JWIntegerDiv, JWImageResizeByLongerSide", and you're using ComfyUI Portable, and you want to apply the fix, DO NOT just right-click-open-in-terminal in the python-embedded folder to do it.
If you do that, it will will update the audio-file-thing in your main python installation, and not the one in the portable comfyui installation.
Use the normal command prompt and navigate to the portable folder, then apply the fix.
@geogy222 Hi - thank you for your hint 👍 and Merry Christmas.
Sometimes I think I should publish a small FAQ in the model description for all the inputs and knowlage here, cause in the comments much is lost after a short time. But maintaining a FAQ would be a lot of work....
@arkinson I mean don't really need to maintain it per se, just add it in the article itself at the bottom and link to the comments? Just something simple will do
@geogy222 Everything changes very fast actually. Comfyui releases a new update every few days. Writing a small FAQ with maybe 10 to 20 entries and links would be relatively easy. The problem would be more to maintain it day by day. "Your" issue is just a couple of weeks old for example - and what today is true could be wrong already tomorrow. On the other hand I often answer to the same questions - more then 100 in the last 8 weeks 🙄🙂 Maybe i will find a simple way for the most asked issues....
Does anyone know how to apply the fix in Linux? Do I have to go to the venv folder (I don't have a python-embedded folder)?
Edit: Found it myself, just activate the venv as always, then type "pip install soundfile", done.
Hello, its me again. I like your workflow, I use it daily, but now I rent a Tesla A100. Can I change GGUF models to full .safetensors? And how if I can?
Sorry, I have no expriances with your gpu and full models like fp16. Generally you can use any model you want. You just have to use the right model loaders and the right clip model. And you might check if you still need lighning loras, BlockSwap, if you can use higher resolutions, frame rates, etc. With some knowlage and a lot of testing this should be no problem.
So far, this is great. Haven't tried the fflf yet, but it works as well as the previous versions. Oddly, the block swap option no longer allows me to adjust block swap and I just have a node I can either enable or disable - a comfy issue for which I am trying to find a solution.
I should add that I loved being able to enable and disable FFLF at will.
@hdean What do you mean with "block swap option no longer allows me to adjust block swap" and "a node I can either enable or disable" - wich node??? Wich comfyui version and release version? You turned Node 2.0 off?? Any node conflicts??
Yeah, the block swap nodes are just blank, now. I haven't looked for conflicts yet. But a few days ago they just decided not to work. (Was using one of your other workfolows). But I am using a 5090 so it really isn't an issue.
By the way, I tried the FFLLF and the lightx loRas were a bit of an unpleasant issue. I switched them out for different lightx loRas and adjusted the strength to 1 and 1 and it worked like a champ.
You do good stuff, man. Wish you had a bigger card and were making flows that used larger models. I switched out the q4 for q8 ggufs for better rez. I really like your work.
Hello gents, by the look of it have been depreciated by comfyui
@ts05961352787 Gotta love it. I tried to put in another block swap but it doesn't want to connect. The node is still affective, I assume under the same settings I was using prior to it no longer "working." I believe it works because when I activate it after an error the WF works fine. Just can't adjust the block swap. Any notion of when or if there will be some sort of fix?
@hdean @ts05961352787 The BlockSwap nodes are no longer necessary. VRAM usage is now managed by the new comfyui core and it seems to work perfect now. You don't need to change anything, because the nodes are simply ignored.
I expect to release a new version shortly with a very interesting Wan22 model and a few minor improvements.
@ts05961352787 Thank you for the hint.
@arkinson My pleasure. I've adjusted a few things. I've a 5090 and can run larger models out works very well fflf.
@hdean If you like, have a look at these models: https://civitai.com/models/2053259?modelVersionId=2540892 I have tested them over the last days and they work perfect, especially with NSFW content, but also SFW seems to work well. You can use the Q8 models, just bypass the lightning loras and be carefull with any additional loras. And if you like to generate longer multiscene videos, have a look at the workflow linked at the model page. You can generate 15 - 20 second multiscene videos in a semi-automatic way with brilliant quality.
@arkinson Thanks much. I will be checking it out in the next few days - if I can get the time, anyhow.
@hdean Or just check my new version v3.0 😉
@arkinson Shit fire, man! You are a busy, busy person! I will check it soon. Be aware, Comfyui has been doing some funky things of late. Among the oddities is that it is not releasing the vram regardless of the vram release node. If it isn't done manually it has been causing the Cuda to fail, and sometimes ends up showing black or static images in your workflow and one other. It also seems to have issues after running a simple zit workflow.
If you get complaints about this, it is a new "feature" in comfyui.
@hdean I know, I struggled a lot for myself with the VRAM issues over the last time. I`m now on latest comfyui v0.7.0. In my case (tested with Flux and Wan22) the new comfyui core VRAM management seems to work very well now (BlockSwapping is not needed anymore). Yes, it seems there is a lot of load into the shared GPU memory, but I get no OOM errors and I do not "feel" any delays in execution. Just test my new workflow 😉🙂 I'm curious to see what bugs I've hidden in it 😅
On the other hand I have a lot of trouble with comfyui on 3 different installations for around 3 weeks:
- I can not drag-and-drop any images into the workflows anymore (not into image loader nodes, not to use the embedded workflows). I allways have to use the File menue ☹️
- Desktop version did not remember my last workflows anymore 😡
- Flux Lora Trainer crashes with "Expecting value: line 1 column 1 (char 0)" 🙄🤬
@arkinson I was trying to load a saved version of your WF into Comfyui the other day and it wouldn't load in at all. I had to grab my stock version and load it, then change to the Q8 ggufs, then save. It works again. So far. Comfy has been a real pain in the ass of late.
I hope to get to testing your newest WF this weekend. I have been finishing up a movie, much of which was generated in your workflow, when Comfy wasn't being a pain in the ass. Now that I am done, I can (hopefully) check it out this weekend. Then again, who knows what the little woman will force on me.
Hey, I do appreciate your hard work. I know what a pain this shit is.
https://www.reddit.com/r/comfyui/comments/1nj9fqo/distorch_20_benchmarked_bandwidth_bottlenecks_and/
tl;dr video is so compute-heavy that it is not affected that much by offloading. That applies to block swap as well as Distorch. So that means everyone with 12 GB can easily handle Q5_K_M, if you have lots of LoRAs you will need higher quants if you want it to converge. expert_mode allocations cuda:0,0.5gb;cpu,*
There are at least two different Wan 2.2 speed LoRAs. idk how they compare to the 2.1 version you are using. Haven't tried >1 strength.
There's a nsfw umt text encoder.
I use FILM VFI as it doesn't flash as much as RIFE does. Since it takes so long I wouldn't have it an auto part of the workflow since you're not sure if you want to keep it. I also use 15 fps as a base so it evenly goes into 60.
You can also use Wan Video Sampler with 12GB if you offload all but one or two blocks, but you will need to use the text bridge node as that can't load the fp16 text encoder. It's Maximum Wonk requiring frequent restarts and VRAM clear nodes but BindWeave and some other Wan derivatives need it.