ELSA3D, a unified 3D generation and captioning model, hits SOTA at half the FLOPs, accepted at NeurIPS
CSProfKGD · x · 2026-10-15
ELSA3D is a new unified 3D model that generates 3D assets and describes them in language within one backbone, led by Tianjiao Yu and accepted at NeurIPS 2026.
- Instead of dumping text + 3D tokens into one flat sequence, sparse Anchor Tokens route each word to the right geometric scale, and an elastic router concentrates cross-modal capacity where semantic–geometric alignment is most needed.
- It reports SOTA on image-to-3D, text-to-3D, and 3D captioning at roughly half the FLOPs and latency of comparable approaches.
- Project page linked in the original post.
More from Multimodal
- HeyGen Voice turns any book into an audiobook via ChatGPT MCP integration — HeyGen · 2026-10-15
- FLUX 3 video upscaler hits Runware with 4K output, Precise and Creative modes — aziz4ai · 2026-10-15
- Reverie ships Intercut Clips to fix AI video's hardest shot: dialogue scenes — lmoroney · 2026-10-15
- Midjourney sref code produces glitter-textured Pokémon starter illustrations — sergeantsref · 2026-10-15
- TRELLIS.2 image-to-3D runs at full 1536 resolution on a 16 GB card via tiling patches — over-yonder-tech · 2026-10-15
- OpenSpline: open-source avatar layer adds a realtime lipsynced face to your voice agent in 5 lines — Mikasa0xdev · 2026-10-15