Now faster and easier to install

This sounds great. Can you explain how you did the auto-cleaning of the RAM and the auto-save for the last frame?

@Fox009 I used a clumsy but effective method.
Add an INT node and connect it to the following two nodes:
1. The location in the original workflow where the length is manually modified
2. Add "🔧 Image From Batch"(fm comfyui_essentials).

The input of the "🔧 Image From Batch" node is connected to "RIFE VFI (I recommend rife47 and rife49)", and its output is connected to "saving the image".

@Fox009 I just found custom node "CLean VRAM used" from "comfyui-easy-use" and add this after VAE
And last frame like @rnvvmw said yeah

Replace "Load & Resize Image" as these step below
1. Add "Load Image Batch(was-node-suite-comfyui)" node
2. Add "Int(comfy-core)" node and connect to index from "Load Image Batch", set value=0; increment
3. Add "Resize Image v2(comfyui-kjnodes)" connect to image from "Load Image Batch", set width=480 ;height=10000 ;upscale=landczos ;proportion=resize ;divisible=16

yes, there is a slight loss of detail, I am very sensitive to that, but I far from agree with that statement, there is a slight loss, but not that significant compared to the artifacts that are present on the original video. Maybe you need to try a different upscaler or lower the denoise of the second pass

Anyway, while there is no magic, there is a way to overcome some problems using available methods, this method is one of the best I have seen, if there are any others, please share with us.

I have a 5090 but I'm using 14b_480p_Q8.gguf and generating at 480x832 (something around it, can't remember the exact number in my head) for 98+ frame in around 5-6min. 39min is just too long for a few second of video. You are better off making small resolution videos, pick the one you like, then upscale it with RIFT or Topaz AI, total time is under 10min.

Just try q4 ver, it takes on my 4070 ti about 10 min for 81 frames

@whiper6993 Yes I might try the Q4, although quality is my main priority so I went for the Q8

@Kristirina Yes it would be interesting to try that method and compare the quality, thanks. By the way, I normally queue around 12 videos and let ComfyUI run overnight, then check out the 12 videos next morning (which have taken about 40 minutes each) so if I was generating videos in 20 minutes I'd have 24 videos to sort through next morning which would be too many LOL

@kennysladefan293 And here is such a topic that you simply will not be able to use the large version normally, not only because of the speed, but because the memory on the processor itself will be overloaded

Sounds good, do you have a json or something for that 2 pass sampler? I'm not sure how to do it lol

Upload this please!

Please more info about this. Thank you!

@HazardAI Awesome, look forward to your future work. Cheers

@HazardAI You can take a look at all the works in my files. Most of them were completed using your methods, which are great. I'm just still researching if there's a better way for smoothing. Have you currently researched any new methods?

@HazardAI nice thanks !

@HazardAI how do i know if wan is fully loaded to vram?

@HazardAI okay thank you ❤️

how long does it take in total on your 3080 10gb to output a video

@bigmantingzz depends on the total size and settings ofc.

but for something like this https://civitai.com/images/76097547 it's around 15-20 mins (for 5-6 sec)
I usually test prompts without upscaling and interpolation, so it goes to just around 5-8 mins; and then turn them on once I'm with the result.

@DollarStoreAbraham thanks thats a great time for 5-6 seconds im gonna try it out !

I get the same error "output with shape [1, 45900, 1536] doesn't match the broadcast shape [2, 45900, 1536]" half the time

@jasonafex Any resolution here?

@Delavestra I believe I ended up switching to a different model/workflow

you might be getting that error if you're trying to generate with a landscaped image instead of a 1x1 or portrait image, to fix it just switch the Width and Height in the MAIN - Load & Resize Image node, and do the same for the PREP VIDEO - Resize Image V2 node.

Additional notes for readers: I am using the Q3_K_M. Instead of the original Wan 2.1 link in page, I use ComfyORG i2v-480p GGUF. I used this since it links up with Civitai. Like when you try to "add a resource used in this gen." Civitai asks before posting.

Additional notes for OP: Also, when you DL v1.2, the zip file gives you two files (one with triton, one with no triton). I'm a some what beginner to ComfyUI and I have no idea how to install this triton thing with pip or python or whatever code wizard thing it wants me to do...so I just use the wan2-upscale-no-triton-v1-2.json file for the workflow. Maybe add a "how to install" this triton thing for clueless people like me if you want people to use it.

Edits: spelling and more notes.

Maybe it is ComfyUI standard node ? I remember it wasn't working on a previous try (so I just disabled it) but since I updated ComfyUi and it seems to work now (but I still disabled it I think)

Right click it and then click Bypass, or CTRL+B

Make sure when you click on a model (inside the nodes) they allow you to select it (meaning the model indeed exist at the right place in your folders) sometimes it could just list the model name while you actually don't have it loaded or not at the right place

Hi, you can try using the flag: --lowvram, or even --medvram, worked for me to counter the mentioned issue. Hope it helps!

@Woolyskink160 hi, I'm having the same issue as well! by using the medvram (lowvram) will this increase the time it takes to load the entire WF?

@Woolyskink160 hi, is it possible to have these VRAM issues on a 5090? I'm using Runpod

I'm using the Q3 versions of the models and a 5sec clip (81 raw frame) fits in my 10GB 3080.
Could give that a try, idk about the more VRAM hungry models tho

This is often an issue with the model you've chosen, are you using the recommended models (Wan, CLIP, VAE, etc?)

it is likely your comfyUi is just outdated, many of the nodes seems native, just update to or reinstall last version

@fouchardmilcoupes311 Possibly. I'm running into the same issue where the rgthree nodes just refuse to install and comfy is up to date.

need umt5_xxl_fp16.safetensors instead

@redsnail got yah. thanks!

for me it does not work well at all? what are you ksampler settings? did you change the shift too?

@testay I just set the CFG to 1, steps to 8, Euler, simple, and removed the teacache and SLG node.

I didn’t change the shift, I only set it to 5

Interesting, so removing Teacache and SLG for causvid, that might explain why my result were retty bad when I tried, i'll give it a try later without teache and SLG

@fouchardmilcoupes311 Basically Teacache and CausVid are punching each other in the face when they are used together in the workflow from what I gathered.

@HazardAI Yes, more testing is definitely needed. I haven’t had any success using your V2V setup using 2 samplers with causevid tho, it always outputs a black image, idk why. That said, for video generation using 2 samplers is generally recommended for improved motion.

As for me, the sweet spot with causevid and your V2V is around 6–8 steps. I have very limited VRAM, so this setup lets me refine a video without waiting for hour. It only takes about 13–15 minutes to refine a 5 second, 1280x720 video.

Maybe you can make another workflow version with causevid.

@HazardAI I've gotten decent results with CausVid as low as 5 steps @ 0.5. For higher quality I went to 8-10 steps with lower lora strength. If using LORAs I usually set the cfg to 2 instead of 1, but still tweaking and experimenting. I also got the suggestion for a 2-sampler approach over at @RalFinger Discord. In my first simple flow I use no CausVid with cfg 6 for the 1st sampler for prompt adherence and CausVid with cfg 1 for the 2nd sampler to refine the output. Even a 4+4 or 3+3 step approach was pretty good. But this is also still in the experiment phase.

@HazardAI But do you get cartoon oversaturated (almost cell shading effect) with CausVid ? because that is what happen to me when I use it, sure it's super fast but too low quality a loss of realism, maybe I'm doing something wrong

@fouchardmilcoupes311 Noticed that too, although mostly when I feed the last frame into the next clip and stitch them together. After 3-4 clips it is rather noticeable. But this latest WAN workflow over at @tremolo28 sounds like a solution: "It contains a Color Match option in postprocessing to cope with the increased saturation, the lora is introducing."

@HazardAI Yeah that might be a good plan B

@HazardAI I put an unload all model after each VAD decode (I mean those after Ksampler of course) seems to work, or at least doesn't seems to bother my PC, tested with Wan 2.1 14B t2v for the refiner too (in gguf format though, seems better quite long though)

type = wan is selected on the clip loader node ? It work for me with Q4 to Q8 gguf version

If all fail, maybe making sure you are using latest comfuUi might solve most of the problems (The number of time just upgrading/reinstalling comfyUI to latest solved problem for me is countless time)

@fouchardmilcoupes311 i did the latest update for comfyui and it kinda crashed everything

@Aiartsenpai you ran the *.bat in the update folder

update_comfyui_and_python_dependencies.bat

this is the real update I'm talking about, if nothing work well maybe make just another install and start fresh, it can't be that bad if it worked before.

@fouchardmilcoupes311 oh i figured it but ty.

64 ram or vram? What GPU do you have?

@HazardAI oh i figured it out. It was a bit buggy. updating fixed it

@HazardAI ty for helping out though

you probably have an outdated 'obsolete' comfyUI version, by not found you mean the nodes ar enot there at all or circled in red like unknown ?

check your memory usage, if you see 98-99M of your V-ram and 98-99% of your system ram used, most likely you are dealing with the program trying to deal with memory problem, it might still complete but you are just not having enough memory, so reduce settings, or make sure you don't run other program at the same time eating your memory, I doubt you will ever be able to make a video in 5 seconds, but 2 hours seems way to high

and by the way, try 4 seconds video, it will significantly reduce any memory usage, maybe the problem is just 5 seconds lenght is a touch too much for your ram

I finally fixed it! It's really the issue of V-ram problem. I tried a node which free up V-ram when loading model, now I can create a 5 sec vid in 5 minutes. Thanks!

@oo4766229 cool, yeah it was really looking like that sort of problem

@fouchardmilcoupes311 Wondering how that happened though. I've seen people with V-ram less than mine maked videos in a reasonable amount of time.

How did you fixed it? I have 12gb VRAm and 16gb RAM

@davifpinto I used this node to load checkpoint https://github.com/pollockjj/ComfyUI-MultiGPU

@oo4766229 Thanks!!

@oo4766229 You figured out some way to use multigpu with i2v?