@METAFILM_Ai I left it at default of one the videos metadata for the steps/frames resolution, I even tried the 480p one and changed the resolution accordingly.

With the built in comfyui method (though text to video) it takes 10 minutes or so (same with hanyuan) I am at a loss as to why its so slow using the workflow of your example videos, not sure what i have set up wrong

sooo , same here, finally made wan2.1 work, hunyuan video on 720p takes like maybe 3 minutes if I just do 20 steps and like 97 frames. I got sage and teacache on tho, but it is insane.
Now ... compared to that
I just downloaded all the models the workflow came set up with, did not change any settings, made an image in flux to 480x824 and right now it seems like this will be about an hour to do. So heeeeelllll naaaaaah wtf, there is no actual way this should take that long. Something ain't right here.

Edit> card is 5090

Edit 2> I am a moron, I kept seeing the size in the promt so figured maybe I can use that size, it worked pretty well on 720x720 , okay so now it is on 6 steps, 1 second 720x380/ish size. So far it is indeed 133 seconds to generate that one second, so I am guessing I can probably bump this thing up to 3,4,5 seconds but have to consider very small size video, probably 720p would not work.

now I am testing it with 480x288 , it took 250 second to do a 5 second long video 20 steps, pretty accurate even without lora. Imma try this with actual realistic size for my card, and tweek some settings, I tink even smaller models could produce pretty good quality to spare some VRAM, I am on 98% with the freaking 5090 , but , the temperature actualld doesn't eve go above 55, so it's a pretty energy efficient workflow and model. Hunyuan with teacache and sageattention can go up to 71 ... okay the average is 65 and the peak temp is 71, but my peak temp with this is literally like 55,56.

Sorry for the novelle.

@sacrificegoat154 the novel is helpful. i will need to revisit this. I assumed the default values found in the meta data of the example files as the ones to use....seems I assumed wrong. Side note, my 3090 is totally not jealous of your 5090...nope...not one bit... :)

@METAFILM_Ai the problem I am having is garbled output using the 720p fp8 i2v. fp8 e4m3fn works fine but not so great quality. I tried a iv 720p Q8 gguf (not sure where I got it from now) which gave worse results. Tons of aliasing and fizzing. I heard there is supposed to be a proper gguf gor 14b 70p i2v based off of bf16 (fp16?). I need an official build of that to test. I am grabbing the Comfy repackaged i2v 14b fp8 model from huggingface at the moment to see if it is the same garbled mess I am getting with kijai's fp8.

@METAFILM_Ai the comfyorg gguf file which i have downloaded where do i put it? I ask because i put it as a diffusion_model but i see that it's a checkpoint ? the problem is that i don't see the results i see in the photos.I want to have an easy time when i want to generate a video where the pussy can be seen easily and not with difficulty

@METAFILM_Ai can you explain more please because the infos under the post are not very informative that's why i would like to know how to load the model is it just with unet loader gguf or do i have to do it other way?

@METAFILM_Ai whats a webp?

@METAFILM_Ai and where is the webp

there is no webp format, only png and drag does'nt work

@METAFILM_Ai nice, i'll test it when it's released from the early access because it looks super lightweight also, i've contributed some buzz. thanks.

@METAFILM_Ai Got it! Thank you so much!

@METAFILM_Ai sorry, I just dropped one of your video files on comfyui for the workflow.

@METAFILM_Ai Does this model have no nsfw？

@KDDKBD its does NSFW!! Thank god!
I was so sick of the stupid censored Minimax and Kling models!!!

No such shit, that would be stupid. You can only take out one frame from a vid and you start with img2vid from that frame

i think you need the T2V model for v2v

To download videos in .webp, make sure to download the videos using the download button at the top right when you open one. If you right click and select save as it will save as an .mp4 in Chrome

@funscripter627 thanks, but it's still mp4

@METAFILM_Ai Hi. Is there any other places to download the webp or json workflow files? Clicking the download button also only gives me a mp4 file like "WanVideo_2_1_00011.mp4". I tried on Edge and Safari.

@1027691161356 I just download a couple of .mp4 files using the download button and they all contain the workflow for comfyui. I downloaded the first couple of vids from the 'Kijai-Wan2.1-I2V-14B-480P' version.

I was originally under the impression that only the .webp files contained the workflow but I was wrong.

@funscripter627 i dont get it , how do i extract the json workflow from a mp4 ?

@funscripter627 but how do i extract the json file from it ?

@hiukiny162 You don't have to extract anything. You can just drag the file from the folder where you saved it into your browser tab where comfyui is running. If it contains a workflow it will load it automatically.

Having the same issue, downloaded all the files and they're all mp4's and tried dragging them all over but they all say Unable to find workflow in "file name" been trying to get comfyui wan working all day but it just doesn't...

sometimes it's work now....

yes, same problem for me 😒

@METAFILM_Ai Nothing downloads as webm? Neither Edge or Chrome will download this as webp. How the heck are you downloading this.

@trashkollector175 If you are using ComfyUI just use the video helper suite to save the video as an mp4. LoRA can also help with NSFW like the titty drop lora for wan or the general NSFW lora for wan.

@trashkollector175 Format doesn't matter, i drag mp4 or webp to comfyui and it opens me a workspace, but some guys are hiding their stuff on purpose and it will/can be empty

@METAFILM_Ai Thanks. It took me some time to fully comprehend 😂

@thiefandliar It's not working for me with mp4 using the portable version. I am running the latest version of comfy. When I select the mp4, it does nothing.. No error message.. nothing.. and when I look at the console there is nothing... If I had an error message of some kind I would provide it.. I converted the mp4 to webp, but then comfy said "it can't find the workflow in the webp", lol. At this point I really don't give a *****

It works with something like the titty drop lora, otherwise it's painful even on batch I would get like 10% success rate. Most people are using complex worklows that don't work or are too slow for my setup for some reason. The original example workflow works at decent speed but the saved file is only webp or WebM for whatever reason. I'm new to this, but I can at least generate something. Let me know if you need help.

you can try my workflow for wan2.1 and test if this helps you https://civitai.com/models/1382864/wan21-image-to-video-workflow

@tawsLoRa I am getting the following error "WanVideoModelLoader" Should never be installed with your workflow. I have no idea how to fix it. Also, I didnt find the place to enter num_frames, i just see the parameter exists but never found it. The resolution is also where? Thanks

@CasualGamers you need to open Comfyui-Manager and then click install missing nodes.

The workflow uses the WanVideoWrapper https://github.com/kijai/ComfyUI-WanVideoWrapper
and if you dont have the Comfyui-Manager https://github.com/ltdrdata/ComfyUI-Manager

also you need the latest version of comfyui

error:
C:\Users\---\Downloads\Compressed\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\custom_nodes\comfyui-videohelpersuite\videohelpersuite\nodes.py:105: RuntimeWarning: invalid value encountered in cast

return tensor_to_int(tensor, 8).astype(np.uint8)

C:\Users\---\Downloads\Compressed\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\comfy\utils.py:825: RuntimeWarning: invalid value encountered in cast

@METAFILM_Ai i upgrade update_comfyui_and_python_dependencies and workflow kjai2v-480 but blank image other workflow it work fine

That should be easier than AMD on windows from what people comment. Generally you need to fiddle with drivers and dependencies a bit, but it's possible, details depend on the exact GPU. Cannot tell you much more, as I chose the pain of getting AMD for Windows and it's been a PITA, but it can be made to work.

I was wondering the same thing... why not just attach the workflow... wth

The other MP4 Videos have a workflow but not the one with the girl in a silver swimsuit. The issues is the workflow also changes for each video.

@Devilday666 but how do i extract the workflow from a mp4 file ?

@hiukiny162 Download the MP4 video and drag and drop it into the workflow.

@Devilday666 i use mimic pc , i can't drag the file into workflow , can someone share a screen cap?

Better get a 24Gb or more graphics card. honestly I have a 3080ti and always running into out of memory.

I run on a graphics enabled EC2 instance on AWS running an A10G

wait for 5080 24gb ver its coming!

I use an Evga RTX 3090 FTW3 Ultra Gaming 24GB GDDR6X, the results are quite spectacular but it's not that fast, I think I have non-optimal settings, I'm trying to fine tune it

On RTX 4000 you get FP8 which if the model uses, you get a decent speed boost. I use a 4070 Super which is similar to what you want, it's got an ok speed, but you probably have to stick to 480p or simpler workflows due to memory constraints. All other workflows I have tried aside from the "stock" one are dogshit slow, so I added loras to the original workflow example from git and mostly using that. I can get 320x240 81 frames in about 10-12 minutes. I do that until i get the prompt right and then run a batch at 640x480 overnight. I am thinking about upscaling later with whatever model is decent. If that seems reasonable to you, the 4070Ti is a decent enough choice.

3070 takes 27 minutes with lora and 15 without lora at 480x976

@NiftyyKnight Can you share your workflow? I was getting way worse results with 3070 so I upgraded to 4070s

@NiftyyKnight How many frames are you generating for those numbers?

Try in my profile, the very first image that looks low/choppy framerate is from a webp file that should have the simplest workflow

if anyone facing the same issue, I solved it using these on the negative prompt: "jewelry, piercing, metallic, glossy, artifact, chrome, shiny artifacts"

Pretty sure the answer is no. (and neither Forge). Looks like ComfyUI is the only option for all things WAN. (Although, I did see YT video about SwarmUI allowign ComfyUI).

I am on the same boat and I think I will eventually migrate over to ComfyUI. It looks overwhelming but I feel like Forge/A1111 are no longer being updated these days.

@Tezozomoctli comfyui is so intuitive and so easy to set up and has so many premade workflow templates that you can literally be up and running in minutes. The only "learning curve" would be the nodes, which you get an understanding of when you try out your first workflow, and the actual "comfyui manager" setup situation, but once you have comfyui manager up and running and are able to update, download nodes, etc in the manager, then its pretty easy from there... and even that's not TOO daunting... its just the first couple of times you try to get something like manager installed you might miss a step or something and it may take a second to figure out, but once you do, you get the eureka moment and everything clicks and its literally the most robust, easy to use, yet, highly customizable AI workflow on earth.

@BreezyHeezy I want to believe that this is the case. But, nearly every workflow that I try to install, that doesn't come embedded in Comfy, seems to have custom nodes that refuse to be recognized.

@lavasplit753 git clone is your best friend, also wget "url here" is your best friend as well. Sometimes, a little coding knowledge goes a very long way in these cases. I suggest, if I may, that if you don't already have some super basic terminal skills (putty, powershell, jupyterlab, etc,) that you get some of those skills. It makes a nearly impossible task (finding and installing custom nodes from outside of comfy manager) into a literal 3 click super fast ftp transfer. And you can even save your git and wget type codes on a cheat sheet in a text file, AND you can lock it as "read only" so you don't screw up the codes accidentally while copying/pasting until you get the hang of it and start remembering them. In A.I. its not "necessary" to have coding experience, but even a little bit will open doors for you. (I'm writing this assuming you don't have much coding experience, but if you do, then you already know what I mean) At that point, when you get your codes straightened out, its literally a matter of typing the node name in duck duck go and finding where its hosted, then getting the direct download link. Wget and git clone will handle the rest. Also, you will need your api token for that type of download from civit, and often you'll have to rename the file and its extension if you use wget to download from civit. If you're running locally, that won't be an issue becuase you can just direct download into your machine. Like I said, I'm typing this under the assumption that you don't know code, but if you do, then you already know what I'm saying. lol

@BreezyHeezy Y'know what? After reading this, I'm going to go for it.

@BreezyHeezy I got comfyui a few weeks ago, and learned the terminal launch and nodes well. my issue is whenever I come to download the required nodes from an online workflow, my comfyui won't always download the nodes completely (there are errors or missing files). what do you recommend?

update the gguf node and try again

I got the issue resolved following this:
https://github.com/1038lab/ComfyUI-OmniGen/issues/37

make sure you have Windows 11 SDK and MSVC build tools installed in visual studio

@ColdBread i did, "AttributeError 209 clipvisionencode"

It seems like you dont have a clip vision model loaded, https://github.com/comfyanonymous/ComfyUI/issues/5559
either set the strength to 0 or load a clip vision model.

@ColdBread i think i got some wrong with model and folder, i dont know , i try to follow it https://comfyanonymous.github.io/ComfyUI_examples/wan/ everything is work

This is the workflow im using, it has a really good guide to where to download everything you need. theres also a simple version if you dont need the upscale or interpolation stuff:

workflow:
https://civitai.com/models/1309369?modelVersionId=1592466
guide:
https://civitai.com/articles/12339

@ColdBread thanks, let me try

@royxyou2303 No problem man, let me know if you have any trouble getting it to work.

@ColdBread is good, i try to use other workflow, all be right, just sometimes i gen video from asian photo, that became a Westerner XD , that fun

im using a 4090 running local and its adding a good 15 to 20 minutes to my gens as well. usual gen time for me is 4 mintes with this checkpoint 19 to 20 minutes

I also suffered until I increased the swap size to about 65GB. As a result, my RAM consumption during the process is ~80 at peak

there are video extend workflows on this site

Wan and other video ckpts aren't supported yet (aside from the handful we currently have). They did mention on stream they were working on/looking to adding Wan on site.

4070TI 12G crashes to blue screen in the middle of operation.

@LiDailin Using which checkpoint?

Yes, I'm using it with a 3070ti 8GB without issues

@admiral_underpants I cannot run 14B 480P, but the GGUF version works.

screw it gonna find out for myself i guess

what video card are you using?

@FEARNIME I'm on a 4090 and I'm also ending up with a pixelated mess of green static. But I'm just now trying to learn to use this. Not sure if I downloaded the wrong model, trying to use the wan video template already in comfyui by default.

@daemondenpublic402 Try spending a few dollars on a runpod with a 24GB 4090 (or greater) so that you can learn how to use it without waiting half an hour to see what you did wrong. Then you can use your local 4090 and get a good result 7 times out of 10. Look up "runpod wan i2v one click deploy".

RTX3060 12GB is my card at the moment. I cant run the 14B one, but the 1.3 one I can.

@METAFILM_Ai 收到，谢谢

I think there's an extension that puts in a 'Comfy'-like tab but you'd probably be better off just using Comfy outright what with updates and such. It's not nearly as daunting as it looks and you get more control over things you'd normally have to leave to Auto/Forge to do. Oh also Forge is superior to Auto1111 in terms of speed of production btw.

Colonel Kurtz ?

I would also like to know if anyone has tested it on Google Collab machines, both in the free version and in the paid version. If anyone can give feedback, I would be very grateful.

https://www.youtube.com/watch?v=0jSDmLtNe_k&t=26s about 14:30 in

@Zeus1111 hero :3

This is a FP8 or what?

@METAFILM_Ai ... wow that was helpful...

Idk if there's been any major change lately, but since WAN came out that has always been a no. Must be ComfyUI.

My Brother, switch to Comfy. i know it looks daunting but its better in the end. more customization and its faster. You can schedule prompts with different loras ahead of time. intrupting is faster. they are no longer going to be updating forge. I did it a month ago and im not going to lie that shit was rough. but i got it and its way better. plus comfy gets all the new toys.

@Azkckr69 from what I understand, Comfy doesn't have a NSFW faceswap node.

@wollyzor it does, reactor (the new one) can be made NFSW with one single change to a file.