FBM1806/minimax-h3-flashgen-4step

I2V: snap Canvas to the uploaded frame's aspect ratio and cover-crop on upload (ports multimodalart/minimax-h3)

Update app.py

Update app.py

Update app.py

Add an 'enhance prompt' checkbox (conditioner LLM prompt rewriting)

README: base_schedule is pre-shift; document the shifted video/audio grids

Apply H3's per-modality sigma shift to the distilled base_schedule (fixes ghosted/superimposed frames)

Document the grid-point semantics behind the 4-step schedule

Restore 4-step inference: pin the distilled sigma grid (num_inference_steps counts grid points)

Document the actual LoRA merge method (merge_lora_ckpt.py + layout transforms)

Fix LoRA merge: de-interleave per-head fused QKV rows before splitting q/k/v

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload requirements.txt with huggingface_hub

Upload folder using huggingface_hub

initial commit