Stream3D: HKUST, MIT, Harvard add memory to 3D generators for stable long video streams
jiqizhixin · x · 2026-08-13
Researchers from HKUST, MIT, and Harvard introduce Stream3D, which adds smart memory to existing 3D generators, selectively caching useful past frames to maintain stability over long video streams without retraining. It outperforms KV-cache reuse and flow-based methods in visual and geometric quality. Paper, project page, and code are available.
More from Multimodal
- Video Generation Models Struggle with Complex Physical Interactions Like Handcuffing — flowersslop · 2026-08-13
- Testing LTX 2.5 Video Generation: A Children's Story Workflow — pfeifits · 2026-08-13
- GPT-4o Demonstrates Impressive 'Pixel Forensics' Capabilities — teortaxesTex · 2026-08-13
- Qualcomm VP on Image Gen Bottlenecks: Separating Scene Planning from Rendering — TWIML AI Podcast · 2026-08-13
- Runway Integrates xAI's Grok Imagine Image 2.0 Model — runwayml · 2026-08-13
- Video Diffusion Models Enable Single-Image Refocusing — CSProfKGD · 2026-08-13