tutorials/video/wan/wan2_2

[giscus[bot]]

tutorials/video/wan/wan2_2

Official User Guide for WAN 2.2 Video Generation workflow example in ComfyUI
https://docs.comfy.org/tutorials/video/wan/wan2_2

[perobueno]

14B takes too long on a 3090. (5 minutes for 1 step) 5B gives an error on the VAE. Any solutions?

[NotDimon]

Almost 4 min for 1 step on 4090. I've changed resolution to 832x480 and number of frames to 73. Now it works like 10 times faster.

[ragtaf]

Using the default 5B TI2V workflow I'm getting some artifacts, particulary often with I2V (T2V is sometimes producing them, but are less noticeable). This artifacts appear like flickering giant pixels all over the generated video. I've only changed the aspect ratio to portrait and lowered framecount so I get quicker gens. Anyone knows if this may cause known trouble or if this issue is common across some configs? Thank you in advance.

[ragtaf]

Just tested with the standard workflow, default, prompt, seed, everything in place and artifacts still appear. Using RTX 4080 16 GB VRAM, 64 GB RAM, pytorch 2.7.1 + cuda 12.8. Also, using sage attention.

[bing420]

Same here.

[jkajkf]

Maybe u guys forgot the light lora

[raffaeleg4]

I'm using a Mac with M3 Chip and is really fast, all goes smooth except those artifacts, giant pixels. I also added a RIFE VFI node for excellent smoother videos. Increasing the steps or modify cfg and sampler/scheduler didn't help at all. I'll keep following up this thread.

[SmithTalksTech]

On the 14B workflows I'm getting an error of:

Requested to load WAN21
0 models unloaded.
loaded partially 128.0 123.0943374633789 0
0%| | 0/10 [00:00<?, ?it/s]

Is this just me running out of VRAM? I have 16 GB

Also noticed on the image to video 14B workflow the Json is loading the FP8 models but the text says to download the FP16 models. I'm a noob so maybe I'm just misunderstanding what this is telling us.

[legarth]

This guide says to use the 2.1 VAE for 14B model... Is that a mistake?

[masterpj555]

supposedly not even in the screenshot it shows it.. im getting errors with 2.2 with that example if i change it.. im testing 2.1 now.. i think something isnt right in the vae of 2.2

[masterpj555]

works on vae 2.1 with that demo.. something is wrong on the 2.2 vae thats for sure

[masterpj555]

i found that the wan 2.1 vae is compatible with 2.2 and most likely a vae for 2.2 was never provided.. the ones you find is likely just a renamed 2.1 vae

[lclichen]

In fact, 2.1 VAE has a smaller compression ratio, while 16x compression of 2.2 VAE results in greater mass loss, so 2.2 VAE is only used for 5B models that require fewer resources.

[kibersektor]

When i tried the 14b i2v workflow using the fp8_e4m3fn_fast i get this error: Exception during fp8 op: CUDA error: CUBLAS_STATUS_NOT_SUPPORTED when calling cublasLtMatmulAlgoGetHeuristic( ltHandle, computeDesc.descriptor(), Adesc.descriptor(), Bdesc.descriptor(), Cdesc.descriptor(), Ddesc.descriptor(), preference.descriptor(), 1, &heuristicResult, &returnedResult)

Switching to fp8_e5m2 made it run however i'm not sure if this will make the results worse, I'm running it on a RTX 4090 and one step is taking roughly 2.3minutes

[masterpj555]

switch to studio drivers on your 4090

[dsphotoblog]

is there any reason that the secon samples is on fixed value?

[comfyui-wiki]

Since the 2nd sample's add_noise is set to disabled, the seed is useless.

[Gunter59]

Ich erhalte beim Workflow folgende Fehlermeldung: KSamplerAdvanced
Given groups=1, weight of size [5120, 36, 1, 2, 2], expected input[1, 32, 31, 90, 160] to have 36 channels, but got 32 channels instead. Alle Modelle wurden neu geladen.

[zg-alexey]

Update to latest nightly ComfyUI, it'll fix the issue.

[franklydegenerate]

The workflows are coming out for me as just an image, not a video. Not sure what's going on.

[franklydegenerate]

edit: Looks like it just appeared that way in the preview. It did make a 5 second long video with my rtx 5090, but goodness it took a while. I'm sure we'll get some speed ups that will make it even better! Thanks for your hard work putting together support on launch day, Comfy. :)

[comfyui-wiki]

I have also added some community resources today, like GGUF and WanVideoWarpper. You can try them if you need.

[fonso2]

Use Lightx2v Lora (I personally use rank 256. There are others that use as low as 64) as well as FastWan, both at 1 strength. Uni_PC, simple, 6 steps, CFG 1, first ksampler ends at step 3, 2nd ksampler starts at step 3. Use a GGUF loader to load Quantstack's quantized models (I use Q5_K_M for both high and low noise on a 4090).

Make the shorter end of your I2V image reference (assuming you aren't doing T2V) 480 pixels, and the appropriate height. I've still been working with 16 fps as from my understanding, only the 5B models were trained at 24 fps. Anyway, do all this and watch your videos generate in under 2 minutes each.

[dsphotoblog]

definetly needs lora speedup

[masterpj555]

what do you use to speed up things with lora?

[comfyui-wiki]

Yeah, I have just added our livestream in the docs. Purz has tried the light2x lora to speed up in the live. You can check if you need. I will also add them to the template recently.

[mikheys]

Error: Given groups=1, weight of size [5120, 36, 1, 2, 2], expected input[1, 32, 31, 160, 90] to have 36 channels, but got 32 channels instead

[zg-alexey]

Update to latest nightly ComfyUI, it'll fix the issue.

[weyn9q]

5b model i2v 960x1280 (!) resolution takes just 3:07 with 20 steps. really fast. it is base fp16. so fp8 should be less then 2min and fp4 just around 1min or even less. with lower resolution ofc faster. i tried lower resolution it takes less then 2min. but quality much worst

[hugalafutro]

the negative prompt in the 5b workflow is in chinese, does it matter? running it through deepl gave me

"Vibrant colors, overexposed, static, blurry details, subtitles, style, artwork, painting, scene, static, overall grayish, poor quality, low quality, JPEG compression artifacts, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, deformed limbs, fused fingers, static scene, cluttered background, three legs, many people in the background, walking backward"

is that what they're supposed to be? or should i leave them in chinese? what about if i want to add more, does the language matter?

[comfyui-wiki]

It's okay. This model works with both Chinese and English

[masterpj555]

Very cool but on the "Wan2.2 14B T2V Text-to-Video Workflow Example" i got a stall with a 4090, 64gb ram, 9950x3d .. Shortening the length gave me an error supposedly caused by EmptyHunyuanLatentVideo.

I lowered the resolution in the hopes it would work but keep getting
"Given groups=1, weight of size [48, 48, 1, 1, 1], expected input[1, 16, 9, 76, 100] to have 48 channels, but got 16 channels instead"

AI told me that there should be a "EmptyWAN2LatentVideo "
But i dont think there is such a thing.

[masterpj555]

Do note: I changed my vae to the 2.2.. i think in there lies the issue.. something aint right with that workflow.. Poster should explain.. perhaps its a work in progress ?

[masterpj555]

works on vae 2.1 with that demo.. something is wrong on the 2.2 vae thats for sure

[mayankagrawal10198]

Can I use wan2.2 with same Causvid Lora ?

[comfyui-wiki]

I think the community has tried that. It should work. Also, you can check the livestream I just added to the docs today. Purz has a demo with light2x lora.

[homewmt]

Can I run with multiple GPUs?

[zhutou748]

这个要怎样解决

[comfyui-wiki]

Currently, ComfyUI's multi-GPU support hasn't been merged yet. If you want to try multi-GPU, you can try this PR

[gitmeAnV]

3090 with 64 ram (no sageattention) works well and pretty fast creating 480x480 @ ComfyUI: v0.3.47 using I2V -14Bfp8.scaled if use WAN21_T2V_14B_lightx2V_cfg_step_distill_lora. Only 6 steps needed.
129frames at 24fps (5s) spent less than 3 min (179s)
Thanks for the Model!

[omgitsgb]

currently training a Lora... and its training lightening fast! kinda crazy! very cool. for 1000 epoch's its only going to take about 15 minutes!

[officefo]

anyone have loras that just stay? even after removing the whole lora loader node?

[Logrui]

Im confused why the workflow tutorial is using the 14B models for Image to Video when the 5B models claims to be able to do both Image to Video and Text to Video

[MasterDenis]

does it work with rocm (linux) on amd rx 9070xt? Just wondering if and how anyone made it work, I Will test too tonight.

[Persite007]

Amazing results with an very old Hp600 (12 cores ... from2007) and RTX3060 12go.
14mn to render 720x720 image to video + faceswap + rife with 61 frames at 30fps. (4s video)
only 5 mn for a 5s video 480x480.
Quality of motions are really awesome.
I can says "The game has changed"

[stasmay]

getting error message "UnboundLocalError: cannot access local variable 'clip_vision_output' where it is not associated with a value" any ideas on what I'm doing wrong? am using the exact workflow

[stasmay]

running flf2v workflow

[comfyui-wiki]

Did you update the ComfyUI before trying the flf2v workflow?

[MOUGE67]

I have this error with FL2V : cannot access local variable 'clip_vision_output' where it is not associated with a value

[MOUGE67]

I have udpate ComfyUI, it's working now

[olafrv]

For those with low VRAM (8GB)...

Start with ComfyUI Official Guide WAN 2.2 Ti2V:
https://docs.comfy.org/tutorials/video/wan/wan2_2#wan2-2-ti2v-5b-hybrid-version-workflow-example

Relevant file locations:

"C:\ComfyUI\models\diffusion_models\wan2.2_ti2v_5B_fp16.safetensors"
"C:\ComfyUI\models\text_encoders\umt5_xxl_fp8_e4m3fn_scaled.safetensors"
"C:\ComfyUI\models\vae\wan2.2_vae.safetensors"

Continue with this tutorial (Minute >=27) works faster:
https://www.youtube.com/watch?v=SVDKYwt-DBg

Use Quanstack quantized GGUF models for WAN 2.2:
https://huggingface.co/collections/QuantStack/wan22-ggufs-6887ec891bdea453a35b95f3

Replace Load Difussion Model node, with Unet Loader (GGUF) to load the GGUF model.
Got a faster ETA e.g. 300s reducing video size e.g. 732x480px.

Relevant file locations:
"C:\ComfyUI\models\diffusion_models\Wan2.2-TI2V-5B-Q8_0.gguf"

[mzugovic-keys]

Anyone come across this issue?
KSamplerAdvanced
Bias is not supported when out_dtype is set to Float32

4090 + 64gb ram. I ran updates across all nodes and have the latest NVIDIA studio drivers

[AntonZhygalov]

wan2.2_i2v_high_noise_14B_fp16
4080 super 16gb VRAM
32gb Ram
64gb paging file

Have this error
The paging file is too small for this operation to complete. (os error 1455).

[comfyui-wiki]

Did you try the fp8_scaled?

[AntonZhygalov]

Didn`t know there was fp8_scaled, i used links here in guide. Just tried fp8_scaled and it works thanks.

[Gruffle666]

Hey I am using the Wan 2.2 workflow with 14b fp8 scaled high and low noise with the 4steps lora lightx2v lora. The speed is impressive but I seem to get brown static and weird visual abnormalities especially when I turn up the quality. I was wondering why this may be happening?

[Gruffle666]

It almost seems like 2 separate videos are playing on top of each other.

[Gruffle666]

I can confirm that it is still happening with a fresh install of comfy ui.

[Gruffle666]

The first 640x640 was fine but as soon as I try and up the resolution I get this problem.

[comfyui-wiki]

I have tested on my end, but I can't reproduce the same issue as you mentioned. Could you open an issue here: https://github.com/comfyanonymous/ComfyUI/issues/

[jkajkf]

It happens too when i run at 832X720 121frames and seems to disappear when i lowered to73frames

[sean78910]

First, thank you so much for the awesome guide and example workflow! 🙏 From the Wan2.2 14B FLF2V Workflow, I had a quick question about the fp8_scaled + 4 steps LoRA part. On the second Ksampler, I see start_at_step = 2 and end_at_step = 10000. Should the end_at_step actually be set to 4 instead, since the other one with 4 steps LoRA uses end_at_step = 4?

[robproctor]

You can set it to 4, but it doesn't make a difference. Steps is what dictates the total steps, so end_at_step 10000 still ends at step 4. It's just set to that so you don't have to change it when adjusting step count.

[sean78910]

Ok..got it and thank you so much!!! Have a great weekend.

[polimorphintobra]

<3
After adding wan2.2_vae.safetensors with wan2.2_ti2v_5B_fp16.safetensors, I get error:
VAELoader
Error(s) in loading state_dict for WanVAE:
size mismatch for encoder.conv1.weight: copying a param with shape torch.Size([160, 12, 3, 3, 3])

Any idea why?
I reloaded wan2.2_vae again, but it didn't help (used your link)

Thank you!

[theIronia]

I found that the generated video will be forced to have a same image to be the first and last frame of the video if I use a >81 value for length in EmptyHunyuanLatentVideo node, and the first frames visibly have higher contrast and saturation.

I can't find out that if this is an expected behavior or not. Has anybody also encountered such behavior?