Meituan LongCat's dNaViT shows discrete codes work as universal visual representations
meituan-longcat · hf · 2026-11-11
Meituan's LongCat team introduces dNaViT, a Discrete Native-resolution Vision Transformer exploring how far a finite discrete vocabulary can represent the visual world, unifying vision and language under next-token prediction.
Method
- Built on a frozen pretrained encoder with lightweight residual modulation and a dual-branch objective injecting reconstruction-aware cues into semantic features
- Grouped product quantization scales discrete capacity without inflating spatial sequence length; a balanced training recipe preserves both semantics and pixel-level detail
Results
- Matches or exceeds task-specific baselines across 12 visual understanding benchmarks, 4 generation benchmarks, and an image editing benchmark
- Acts as a versatile visual tokenizer supporting unified encoding alongside high-fidelity image decoding
Project page, GitHub, and Hugging Face weights are open.
More from Multimodal
- Fable 5.2 nails four pro design tasks: 689 robot parts, 4,871 plumbing plans — FinanceYF5 · 2026-11-11
- Fable 5.2 nails four pro design tasks: 689 robot parts, 4,871 plumbing plans — FinanceYF5 · 2026-11-11
- Dev Rebuilds Demis Hassabis' 1994 Theme Park Game One-Shot With Gemini 4 Carbon in 36 Minutes — DeryaTR_ · 2026-11-11
- Designer blends Blender MCP into 3D brand concept workflow for olive oil — sidahuj · 2026-11-11
- Solo dev ships a working iPhone/iPad pickleball game with Claude Code + Unreal Engine 5.9, sharing the full AI toolchain — chongdashu · 2026-11-11
- Hand-painted game built in under an hour with Fable 5.2: 1.5M brush strokes per frame at 200fps in a browser — Scobleizer · 2026-11-11