ali-vilab
How many parameters are actually trained in VACE？
May I ask how many parameters VACE actually involves in training? Also, I noticed that the Hugging Face library provides a copy of the original WAN weights—are these exactly the same as the pre-trained WAN model? Why isn't VACE saved separately as a distinct component? I'm quite confused. Thank you for your response.
The VACE model has only undergone Context Adapter Tuning and has not made any parameter changes to the original Wan2.1-T2V-1.3B/14B.
This design converges more quickly without compromising the capabilities of the original base model, which is beneficial for all subsequent plugin-type functionalities based on the original T2V model within the community.
The Hugging Face library (Wan2.1-VACE-14B) shows that the loaded model consists of [original T2V parameters(part 00001~00006) + VACE module parameters (diffusion_pytorch_model-00007-of-00007.safetensors)], mainly for convenience in loading and code reusability considerations.
