Marigold V2 Rethinks DiT for Depth Estimation, 16-26% AbsRel Gains on KITTI/ETH3D
Cohere_Labs · x · 2026-10-08
Cohere Labs' Open Science Community announced next week's sessions (Oct 12-16), headlined by Marigold V2: repurposing DiT-based image generation models into monocular depth estimators via single-step inference from pretrained flow-matching models, with a 2-stage fine-tuning protocol and a novel Sinkhorn-based loss. It improves AbsRel by 16-26% over prior best on KITTI and ETH3D, resolving fur, foliage, and hair-thin edges, and transfers to surface normals and intrinsic decomposition. A second session covers Semigroup-JEPA for zero-shot physics generalization.
More from Multimodal
- Reddit user shoots a cinematic short of H.G. Wells' The Cone with MiniMax H3 — nikhilprasanth · 2026-10-08
- Ideogram ships 4.5 image model with new app UI and video integration — bennash · 2026-10-08
- Dreamina launches Dreamina Originals, first content label for AI films and series — LudovicCreator · 2026-10-08
- Reverse-Engineer Any Video into a Seedance Prompt with Claude or Gemini — TawohAwa · 2026-10-08
- MiniMax demos code-driven video: text to JS animation and explainer videos — Hailuo_AI · 2026-10-08
- Turkish TTS trained from scratch with the Drifting method on a single RTX 5090 — kadir_nar · 2026-10-08