gdhe17

/

Self-Forcing

like 47

We only use 1.3B generators in this work and the 14B model used for distillation (i.e., the real score or teacher model) is just the official Wan14B checkpoint (https://huggingface.co/Wan-AI/Wan2.1-T2V-14B)

We only use 1.3B generators in this work and the 14B model used for distillation (i.e., the real score or teacher model) is just the official Wan14B checkpoint (https://huggingface.co/Wan-AI/Wan2.1-T2V-14B)

Oh, I meant to ask is 14B going to be distilled like https://huggingface.co/lightx2v/Wan2.1-T2V-14B-CausVid so we can generate much faster using the Self-Forcing method in an autoregressive way, with the causvid model we can generate videos in 9 steps, so I was hoping for something like that for the 14B model so we could generate much faster. Basically a 14B Self-Forcing model.

Basically a 14B Self-Forcing model

I agree with @natalie5 and would be interested to see a 14B version of Self-Forcing as many folks are probably using 14B CausVid or the more recent 14B FuxioniX and FusionX GGUF merges.

I'd def try this out if there were a 14B available. Maybe @BenjiAI would do a YT video demo like FusioniX if it works well!?

Thanks!