Stream3D turns frozen 3D generators into streaming models with bounded memory
pliang279 · x · 2026-07-29
Stream3D is a CVPR workshop acceptance for long-horizon 3D asset generation from monocular streams. The framework turns frozen 3D generators into streaming generators using evidential memory, so it can keep geometry and appearance more consistent across long sequences without retraining.
Key claims include a constant memory footprint, better long-stream consistency, and a training-free wrapper that scores which past views matter most for each 3D token. The project page also reports quantitative gains on GSO and NAVI.
- Training-free wrapper around frozen single-view 3D generators
- Bounded cross-chunk memory, independent of stream length
- Evidence-aware selection of historical views for each token
More from Multimodal
- Developer turns ChatGPT Real-time Voice into a desktop 3D persona — BLUECOW009 · 2026-07-29
- Qwen Audio 3.0 Realtime Plus costs $4.42 per input-audio hour in testing — ArtificialAnlys · 2026-07-29
- Qwen Audio 3.0 Realtime Plus tops Speech-to-Speech benchmark at 84.1% — ArtificialAnlys · 2026-07-29
- AI black-comedy short film turns a zombie apocalypse into a music video — Aggravating-Chest997 · 2026-07-29
- Claude 5 Opus generates a textureless dirt-road car demo entirely on its own — ChrisGPT · 2026-07-29
- TILT improves compositional text-to-image generation with a model-intrinsic reward — Debottam Dutta · 2026-07-29