Testing only: these are H3 right now, despite the repo name. do not merge, train on, or distribute. Outputs can be shared, and written opinions and info about the model can be shared, but not the file.
Beta6 important info:
Built on fast-vsa 8 step. First version is fl2va for t2v and i2v only, reference/hybrid version will be different. Does not play well on alternate samplers. res_multistep is fine though on FAST for audio and motion on this version. Use the comfy template FAST video H3 workflows with their sparse node setting and 10/3 shift, the mix is balanced exactly around those settings. Strength has a magnitude ceiling, tokenizer layers were excluded (fully base model prompting), and the vast majority of lora noise was cutoff by a magnitude floor which allowed the 30+ loras to contribute without any normalization and also without destroying the model. When stacking loras on H3 that is what causes the visual and audio degrade: late blocks heavy lora noise that stacks and unwinds guidance.
