Self-Forcing / CausVid / Accvid Lora, massive speed up for Wan2.1 made by Kijai

2.2

I2V versions are out, and T2V versions got updated.

Here is a good 3 Stage workflow that works well with the lightning loras, minimizing their effect on motion: https://files.catbox.moe/bfgbof.json

Here is a custom sampler made for wan2.2 that correctly changes at the correct timesteps. It requires you have the ram to load both models at the same time though (64GB+), should help quality if you want to use that instead.

https://github.com/stduhpf/ComfyUI-WanMoeKSampler

2.1

7/15 update: I added the new I2V lora, seems to have much better motion than using the old text to video lora on a image to video model. Example is 4 steps, 1 CFG, LCM sampler, 8 shift. I uploaded the new version of the T2V one also.

I'm also putting up the rank 128 versions extracted by Kijai, they are double the size but are slightly better quality.

I suggest using it with the Pusa V1 lora as well, it seems to improve movement even more: https://huggingface.co/Kijai/WanVideo_comfy/tree/main/Pusa

No need for 2 sampler WF anymore IMO. Just plug it into your normal WF with 1 CFG and 4 steps or so. Could prob sharpen it with another pass if you wanted but it no longer hurts the movement like before for image to video.

Full Image to video Lightx2V model: https://huggingface.co/lightx2v/Wan2.1-I2V-14B-480P-StepDistill-CfgDistill-Lightx2v/tree/main/distill_models

Old:
lightx2v made a 14B self forcing model that is a massive improvement compared to Causvid / Accvid. Kijai extracted it as a lora. Example above was generated in about 35 seconds on a 4090 using 4 steps, lcm, 1 cfg, 8 shift, still playing with settings to see what is best.

Just use the 14B self-forcing one, the rest are terrible in comparison now. And please don't send me buzz or anything, give the lightx2v team or kijai support if anyone.