Sign In

Wan 2.2 txt2vid 3 pass test, oral insertion i2v lora

Loading Images

Here's my secret sauce: 14 steps, 1 ksampler does 3 steps on Wan 2.2 high noise (no lightning lora) Euler/Simple, the second ksampler does 9 steps on Wan 2.2 low noise (no lightning lora), the final ksampler does 2 steps on Wan 2.2 low noise with lightning lora. You can adjust the initial high/low pass slightly but 3-4 steps seems to be the sweet spot for high pass.

NO easycache/teacache/magcache. Sometimes I use dpm++2m or res_2s instead of euler.

In my testing I cannot get a better result even when using 20-30+ steps of Wan 2.2 high+low noise without lora. The motion does not seem to be hurt by all, the lightning lora for the final low pass just being an extra bit of denoising.

On my 4090 I could generate this at 640x640 at 49 frames in only 200 seconds. I'm about to upload an example video on my page now, for anyone interested. Upscaled with 4x-ClearReality then downscaled 0.5x

Comments