lightx2v

/

Wan2.2-Lightning

like 221

Hi, could you please provide a workflow so that we can find why the generation quality is not as expected.

Before releasing the wan2.2 lightning weights we have compared comprehensively with the wan2.1 T2V weights (lightx2v/Wan2.1-T2V-14B-480P-StepDistill-CfgDistill-Lightx2v) on a test set and find an improvement by a large margin.

You can start with our workflow https://github.com/ModelTC/Wan2.2-Lightning/issues/3 and change the prompt to check the generation results

You can use the Wan 2.2 14B workflow provided by the official ComfyUI template and add Wan2.2-Lightning Lora for testing.

The resulting video is grayish and has poor motion.
You must increase the Lora weight to 2, but the color granularity is still inferior to Wan2.1-Lightx2v Lora.

High and low noise （weight=1.0）, cfg=1.0, step=5 .

I just checked the issues, and you provided WanVideoWrapper.
This may differ from the ComfyUI official template's integration with Lora. However, the results are similar to those reported in the issues.
https://github.com/ModelTC/Wan2.2-Lightning/issues/3

However, using Wan2.1-Lightx2v Lora on Wan2.2, everything seems to work fine. The only issue is a lot of weight warnings in the console.
I think you can fix the weight warnings in Lightx2v Lora before the new Lightning is released, as most people are still using the original Lightx2v Lora on Wan2.2.

Hi, after testing, I can also say that although there are errors in some keys when loading the WAN2.1m LoRa, they do work on the WAN2.2 models! The LoRa does work, and the result is much better than the new LoRa specifically for WAN2.2. I've tested it with ComfyUI and PyTorch in the console. NOTE; Image To Video Test (I2V)

With Loras new wan 2.2:

With Loras from Wan2.1 on wan 2.2:

have the same problem. 2.2 lightning specially in the high noise model kills all the complex motions.. people are messed up. cinematic feels is gone. etc. it's better to use the 2.1 lightx2v lora on the 2.2 even.

Hi, could you please provide a workflow so that we can find why the generation quality is not as expected.

Before releasing the wan2.2 lightning weights we have compared comprehensively with the wan2.1 T2V weights (lightx2v/Wan2.1-T2V-14B-480P-StepDistill-CfgDistill-Lightx2v) on a test set and find an improvement by a large margin.

You can start with our workflow https://github.com/ModelTC/Wan2.2-Lightning/issues/3 and change the prompt to check the generation results

Your provided examples look really bad. You say it's good just because it looks sharp doesn't mean it's good. Motion is completely messed up. Hope you can see it now.

@acaimari @AgustinCaniglia

Hi, I can't find workflow inside the video that you uploaded.
May I get a workflow to check why it's not working correctly?

One possible reason may be you are using I2V workflow with the T2V loras, this will destroy the generation.

We just release I2V loras ( https://huggingface.co/lightx2v/Wan2.2-Lightning/tree/main/Wan2.2-I2V-A14B-4steps-lora-rank64-Seko-V1 ), please take a try. If the result is not good, please send me the workflow so that I can find out the drawback of our models.

Hope this response helps :)

Is the v1.1 wan 2.2. lightning lora not compatible with the rgthree lora loader?

When I try to use it with the T2V 14B Wan 2.2 fp8_scaled model, I get a ridiculous number of 'lora key not loaded' errors in the comfy console.

@gabrielconroy

We have not tried rgthree lora loader, please try this workflow https://huggingface.co/lightx2v/Wan2.2-Lightning/blob/main/Wan2.2-T2V-A14B-4steps-lora-rank64-Seko-V1.1/Wan2.2-T2V-A14B-4steps-lora-rank64-Seko-V1.1-forKJ.json for T2V.

Hey all. I've had the same troubles as others. Previous loras seem to work much better. The new ones are giving a lot of odd blurriness. Not sure if I'm doing something wrong or what.

use new v1.1 loras at 1 strength and old v1 lora from kijai at 1

Hey all. I've had the same troubles as others. Previous loras seem to work much better. The new ones are giving a lot of odd blurriness. Not sure if I'm doing something wrong or what.

Hi, thank you for your feedback. Could you please provide a workflow so that we can help to find it out.

Also getting poor results, blurry, grainy, bad quality.
With the native comfyui template/workflow.

We are working on the native comfyUI workflow now, thank you for your feedback. @sdfasdfasdfsdfsdf

You may take a look of the Kijai's wrapper - compatible worflow before we get it done.

There is a perfectly functional and optimized native implementation already in comfyui.
And thats what I and many other people will be using.

Kijai's wrapper is bloated and uses too much memory. I cant even do 1 step at 480p, 61 frames without OOM. It instantly goes over 16gb vram.
With native comfyui nodes, I can do many many steps at 480p and full 121 frames - its super quick, super optimized and fits into 16gb only using ~14.5gb of vram in total.

@sdfasdfasdfsdfsdf @gabrielconroy

We update the weights and there should be no loading errors now.

We have released the template workflow for both T2V and I2V. Please refer to the discussion. The workflow file that contains "forKJ" is for kijai's wrapper and that contains "NativeComfy" is for native comfyUI.

@X-niper great, thanks! Downloading now.

I used Wan2.2-I2V-A14B-4steps-lora-rank64-Seko-V1-NativeComfy.json for this fix.

Now, it actually works for the official ComfyUI model workflow, eliminating warnings and improving image quality.

However, a new problem has arisen: after loading Wan2.2-Lightning Lora, the generated video has less character and camera movement than the original model, sometimes even no movement, and slow motion often occurs.

However, a new problem has arisen: after loading Wan2.2-Lightning Lora, the generated video has less character and camera movement than the original model, sometimes even no movement, and slow motion often occurs.

Try using 1.5 strength on high lora, and 1 on low lora, best results likely steps with 6.