MosaiChunk paper proposes spatio-temporal memory for consistent video, tested with MiniMax H3
smereces · reddit · 2026-10-19
A Reddit user spotlighted MosaiChunk, a research project on compositing spatio-temporal memory to preserve character and scene consistency in long video generation (project page: mosaichunk.github.io). The paper reportedly uses MiniMax H3, and the poster asks whether the method could be ported into ComfyUI workflows for MiniMax H3 text-to-video and image-to-video generation to fix consistency issues.
More from Multimodal
- Microsoft MAI reveals in-house 3T-parameter base models with native encoder-free VLM pretraining — burny_tech · 2026-10-19
- Dark Souls-Style Knight Art 'Find the Strength to Carry On' Hits Reddit — Sufficient-Ice8225 · 2026-10-19
- Creator Combines Midjourney and Seedance 2.5 for a Surprise AI Short — OVolosin82152 · 2026-10-19
- Why does 480p Grok Imagine video look cleaner than 1080p open-source Wan/H3? — yolsuzyolcu · 2026-10-19
- Running Moss image model on a 4GB RTX 3050, optimized down to 62s per image — sanketshinde2004 · 2026-10-19
- Reddit asks: why is open source still so far behind Seedance 2.5 for video? — ByteSize_Chaos · 2026-10-19