GeoFix fixes 4D generation geometry with a single calibrated reward, +39% WorldScore 3D consistency
Yeobin Hong · hf · 2026-10-14
Researchers from KAIST, ETH Zürich, Google and EPFL propose GeoFix, a unified framework improving geometric consistency and visual quality when video diffusion models serve as priors for 4D generation.
- Uses a single geometric reward — cross-view depth consistency in static scene regions — avoiding the texture degradation of multi-reward setups
- Key insight: depth agreement rises with camera field-of-view overlap even in real videos, so smaller camera movement can masquerade as better geometry; the reward is calibrated against this overlap-dependent trend
- Applicable to both post-training and inference-time reward alignment
GeoFix achieves a 39% improvement in WorldScore 3D consistency and a 4% gain in VBench Quality Score.
More from Multimodal
- PixAI open-sources Tagger 1.0, the 30,877-tag anime tagger behind Tsubaki.3's data curation — bdsqlsz · 2026-10-23
- Gemini generates a polished character design sheet for a painter badger — azed_ai · 2026-10-23
- NeurIPS Spotlight KoalaGS allocates Gaussians by image complexity for smooth LoD rendering — zhenjun_zhao · 2026-10-23
- Lovart plugin goes live in Codex: generate brand visuals and ads where you code — OpenAIDevs · 2026-10-23
- Same prompt, 2.4x cost gap: AI UGC videos at $0.75 vs $1.78 — miilesus · 2026-10-23
- Open-source Phosphene renders Hailuo H3 2.3x faster locally on Mac — cocktailpeanut · 2026-10-23