Wan2.2 5B Turbo; How Is It 4 Steps?
For some reason, no matter what samplers I choose, no matter what CFG and loras I change or disable; I cannot get any solid results like the examples shown without like 15-25 steps.
What in the world am I doing wrong? I'm just using a generic Wan2.2 5B workflow from Civit, very simple and straight forward.
I didn't test it much, but I did notice it didn't seem to work great with T2V.
With I2V I did get results like this at 4 steps, cfg 1.0:
Yes with I2V I see that it does produce solid results at 4 steps. I wonder what the issue is with making a 4 step t2v work. Thanks for your response!
Idk why but i found this funny.
I saw that the author of WAN 2.2 Turbo opened an issue on FastWan repo, asking why it seems doesnt work for I2V (https://github.com/hao-ai-lab/FastVideo/issues/711). The FastWAN author replied that FastWAN was only trained on T2V data, so it might not work well with I2V.
Then, a few days ago, he released WAN 2.2 Turbo, which was trained on I2V data, and after trying the Turbo lora for I2V, it definitely performs much better than FastWAN.
Honestly, I don't fully understand how the training process works, I used to assume it was all the same since it’s the same model.