DreamWorld: Geometry-Grounded Two-Stage Video Diffusion for 3D-Consistent World Modeling (ECCV 2026)
机器之心 · wechat · 2026-08-19
HiDream.ai's DreamWorld, accepted to ECCV 2026, tackles unreasonable geometry, object distortion and cross-view inconsistency in camera-controlled video diffusion under large viewpoint changes. It injects spatial priors from a pretrained 3D foundation model and adopts a Geometry-then-Appearance pipeline: a Geometry Video Diffusion completes target-view geometry features from a single image plus camera trajectory via feature-level flow matching, then an Appearance Video Diffusion conditions on those features to synthesize the RGB video.
This geometry-appearance decoupling lets structure reasoning and photorealistic rendering specialize separately. DreamWorld leads Novel View Synthesis on RealEstate10K and Tanks-and-Temples, and scores 75.04 average on WorldScore with stronger 3D consistency.
More from Multimodal
- FLUX 3 Upscaler goes live on Higgsfield: 50% faster, cheaper, native 2K/4K video — robrombach · 2026-10-16
- Two Midjourney --sref style codes that nail painterly fashion illustration — sergeantsref · 2026-10-16
- MLP-Splatting (ECCV 2026) code released: MLPs as light-field primitives instead of Gaussians — AjdDavison · 2026-10-16
- Digital human Fae codes her own facial expressions frame-by-frame with Claude Opus 5.5 and FACS — repligate · 2026-10-16
- Meta's Omnilingual ASR, covering 1,600 languages, named a TIME Best Invention of 2026 — AIatMeta · 2026-10-16
- ATAS: Zhejiang-led team builds world-state memory framework for consistent agentic video generation — CSProfKGD · 2026-10-16