You don't have to pay attention to my idea.
In case this could be a new idea for your Wan2.2: if you want it to generate audio, there is a way to do it, but I don't know if it really works. What I can think of is extracting mova-720p/360p from openmoss-team as a patch + custom code.
I saw your comment on wan-dancer and immediately searched for it, downloading the full model as we speak. It’s very interesting
too much is happening in AI, and we always miss good stuff 😄 . It looks very promising. I’ll test it out and see what we can do with it. If I come up with anything remotely interesting, I’ll share it 😁 . By the way, thanks for the tip.
I was just wondering if it's possible to plug the img_emb layer or face_adapter components (fuser_blocks, face_encoder) back into Wan 2.1 I2V?
I was just wondering if it's possible to plug the img_emb layer or face_adapter components (fuser_blocks, face_encoder) back into Wan 2.1 I2V?
Anything that wasn’t trained together usually won’t work properly. grafting modules can sometimes work, but it often introduces new issues you didn’t have before (strong identity won't work good with prompt: https://huggingface.co/rzgar/Wan2.2-I2V-Weak-ID).
You mean porting it into Wan2.2? Wan2.1 already has img_emb. I tried a similar idea here:
https://huggingface.co/rzgar/r_drive/tree/main/W2.3_B
It’s basically Wan2.2 I2V + Wan2.1’s img_emb + LoRA compatibility layers.
Let’s do a test: which clip is Wan2.1, Wan2.2, and Wan2.2 with grafted Wan2.1 img_emb + LoRA keys (Wan2.3_B)?