tutorials/video/wan/wan2_2

[giscus[bot]]

tutorials/video/wan/wan2_2

Official usage guide for Alibaba Cloud Tongyi Wanxiang 2.2 video generation model in ComfyUI

https://docs.comfy.org/tutorials/video/wan/wan2_2

[jancovic]

hello
i was testing Wan2.2 TI2V 5B Hybrid Version Workflow Example form this page,
image to Video not working for me, basicly loaded image was ignored,
Comfy running in Linux latest ver. 0.60, ctrl- B to activate was not doing anything, however lines were connected full blue limage line.. so i think not the issue. No bypass i found in settings of node.

[jancovic]

ok, Tested again, redeployed the workflow and is workign !

[ARDEACT]

Works well but how do you extend the video? Wan 2.2 has a dedicated video extend section, I wonder if it's possible to do one here too

[gershu-ar]

I suggest getting the example updated template for wan 2.2 from inside of ComfyUI. In that example you can set everything: length, size, FPS, etc. The workflows here are outdated.

[0Eagle]

video_wan2_2_14B_i2v.json :No link found in parent graph for id [129:85] slot [7] cfg

[gershu-ar]

Check this:
Comfy-Org/ComfyUI#13189 (reply in thread)

Here is how you access the subgraph:
Comfy-Org/ComfyUI#13189 (reply in thread)

[skylandanimst-star]

Wan 2.2 image to video is broked((( Error
Cannot read properties of undefined (reading 'output')

[gershu-ar]

If you upload your workflow, I can review it and see what is wrong.

[andreszs]

With the Wan22ImageToVideoLatent node, the KSampler's preview doesn't work. It just shows a static, heavily blurred version of the input image all along the sampling process.

[gershu-ar]

As far as I know, there are no working nodes or mods for video preview (I could be wrong). I use WAN practically every day and haven’t found one yet. Why? Latent safetensor sequences are not displayed in the KSampler preview, which is natively built for showing only a single render/picture (one frame, 2D).

Can it be done? I don’t think so, at least with how ComfyUI is currently coded. The preview shows the process of intermediate denoising, which is not the same as showing the final decoded render (a fully finished image).

There are alternatives you could code using the save batch → ffmpeg → gif/mp4 path. But that would consume valuable rendering resources just for the preview. Is it worth it?

Put it this way: if your rendering times are long enough that you feel you need a preview to avoid wasting time, then by definition you don’t have the resources to generate a real-time video preview. Previewing would practically double the rendering time. Why preview one video when you could just render two in the same amount of time? :)

[andreszs]

Thanks, I definitely don't have the resources for that with my RTX 3060 12 GB, a mediocre card by today standards. It's barely decent enough for static images.

[gershu-ar]

A 3060 is still a good GPU for image rendering (those 12 GB are worth gold nowadays). Nonetheless with Ampere (30xx) you will most def struggle with videos, they are demanding! I've got a 3090 and it takes me around 120 to 150 secs to render a 512x512 vid at 20 FPS (Wan 2.2, 14B, fp8_scaled, render + refiner, 4 loras, 40 steps, euler a), so I feel ya.

My suggestions:

• Try other attention methods (psytorch in video can get bulky). Try Sage and Flash attentions and evaluate the it/s. Won't make any change in images but it does make a huge change in videos.

Quick guide: https://diffusiondoodles.substack.com/p/demystifying-attention-in-comfyui
Download sources: https://github.com/wildminder/AI-windows-whl?tab=readme-ov-file#available-wheels

• Never go over 1 in batch.
• Render in lower resolutions: 512x512, 640x480, 480x640, 480x480 and so on.
• Sometimes using --disable-cuda-malloc can avoid VRAM fragmentation in lower mem GPUs
• Beware of Windows' swap file management in case VRAM and RAM are not enough; keep memory at bay
• Try with fp8_e4m3fn encoding, use fp8_e4m3fn models and weights. fp8 uses half the VRAM but can introduce artifacts if all the components are not aligned.
• Test, test and keep testing until you find the sweet spot

Cheers!

[Prabhjot2s]

Hello! I’m relatively new to AI video generation and could use some optimization advice. I’m running an RTX 3090 Ti, but while others are seeing 2-3 minute render times, my generations are taking over 10 minutes.

I am using the Wan 2.1 wrapper for the Long-Cat (16GB VRAM) model. I've noticed my VRAM is maxed out and my system RAM usage is hitting 75%. I suspect the bottleneck might be the VAE Distill LoRA I downloaded separately.

Could you share your specific Wan 2.1 template/workflow? Also, if you have any recommendations for Discord channels or websites for troubleshooting these specific builds, I’d really appreciate it!

[gershu-ar]

Hey, amigo. Long reply ahead.
With a 3090 Ti -Ampere top of the line- your timings should be more than decent.

I've been toying with image to video on Wan 2.2 14b fp8 (based on official template) and integrated it RIFE interpolation (deep learning-based method for video frame interpolation, capable of generating intermediate frames between two given frames in real time).

Visual results are fantastic, no artifacts, no weird stuff. Just keep in mind to use good prompting: Wan is very sensible to it same as video resolutions; calculations get exponential the more you increase it. It's all about math balance!

Find the workflow attached:
Image2Video Wan2.2 plus RIFE interpolation.json

Also the example image:

Have been running it on a 3090 with fantastic results at an average of 46 s/it; at 512x512 pixels output goes down to 26s/it.

I suck at making workflows, but the idea is there:

• Image to video, use good prompting
• RIFE interpolation integrated (Node Initial FPS indicate rendering FPS -the hard work-, node Final FPS indicates interpolated final video FPS -super fast process-).
• Length on node WanImageToVideo is achieved by FPS * video seconds. Example: 15 FPS at 10 seconds, it's 150. Set 150 (or near above) for a 10 second video generated at 15 FPS. If you set 20 on Final FPS, RIFE interpolation will super fast increase video fluidity by 5 FPS.

Also, as a suggestion, try running ComfyUI with Sage Attention and making sure it uses fp8 encoding by loading with switches:
--fp8_e4m3fn-text-enc --use-sage-attention

In my setup Sage attn makes better use of memory and render speed at the expense of a little loss quality; negligent to none in most cases, test it, it makes about 25 to 30% lower rendering times difference on my 3090 than just using default pytorch/xformers or Flash Attention.