Mushishi Creative Stack shipped

Local AI video production: generation, object removal, masked edits, and 4K60 finishing, with a forensic-analysis bridge that starves diffusion models of room to hallucinate.

github.com/MushiSenpai/mushishi-creative-stack ↗  ·  ComfyUI · FLUX.2 · Wan 2.2 · HunyuanVideo · SAM3 · SeedVR2 · RIFE


Six named, tested workflows — from seed-locked fast iteration (⚡ Flashfire) to object removal that takes the object and its reflection (🫥 Vanisher). The interesting engineering is upstream: a multimodal “forensic mode” produces dense scene constraints before any diffusion model runs, so generators obey specifications instead of inventing shadows.

Quality benchmarking against client-grade standards is happening in public — the benchmark table ships with empty cells and fills in weekly, pass or fail.

Posts about this project