license: other
license_name: minimax-h3-community-license-agreement
license_link: https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE
pipeline_tag: image-text-to-video
base_model:
- MiniMaxAI/MiniMax-H3
tags:
- text-to-video
- image-text-to-video
- image-to-video
This is a lora for the video h3 model, using wan2.2 attn layers on the mid to late H3 blocks with peak mapping. Wan's blocks are reversed to align better. Wan is 16fps but again H3's design seperation doesn't make the much of an issue, timestep isn't adjusted at all.
This is extraction result from Wan2.2 branch of my attn graft experiments. At rank 512 the lora only captures 41.6% of the residual orthagonal delta inside the merged full model version for test1, 1024 captures 69%. Both produce a large shift in motion style in terms of prompt response, tending towards old Wan style repetetive motions. The Wan model used is subjective and I'm still looking for reccomendations on better ones (fp16 needed). All experimental. Refer to Eros Max page for exact methodology of how cross-model attn grafting works inside H3 model.