Research map / Vision-language and multimodal / Human and embodied modeling
OneCAT: Decoder-Only Auto-Regressive Model for Unified Understanding and Generation
Han Li, Xinyu Peng, Yaoming Wang, Zelin Peng, Xin Chen, Rongxiang Weng, Jingang Wang, Xunliang Cai, Wenrui Dai, Hongkai Xiong
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add distillation of all layers' hidden states from the teacher (ablation) | OneCAT-1.5B understanding stage · 7-benchmark average (8B tokens) | average score: 31.4 -> 35.3 | Improved |
| distill all layers instead of only the last layer's hidden states (ablation) | OneCAT-1.5B understanding stage · 7-benchmark average | average score: 33.9 -> 35.3 | Improved |
| use full-layer distillation instead of EvE or VoRA distillation (ablation) | OneCAT-1.5B understanding stage · 7-benchmark average | average score: 33.4 (EvE), 32.3 (VoRA) -> 35.3 | Improved |
| use custom teacher (InternViT + same Qwen2.5 LLM) instead of Qwen2.5-VL teacher (ablation) | OneCAT-1.5B distillation · 7-benchmark average | average score: 33.7 -> 35.3 | Improved |
| distill both visual and text tokens instead of visual tokens only (ablation) | OneCAT-1.5B distillation · 7-benchmark average | average score: 34.8 -> 35.3 | Improved |
| add scale-aware adapter (SAA) (ablation) | OneCAT-1.5B stage-2 · GenEval | GenEval: 78.1 -> 81.2 | Improved |
| add scale-aware adapter (SAA) (ablation) | OneCAT-1.5B stage-2 · DPG-Bench | DPG: 74.0 -> 74.9 | Improved |
| increase generation tokens in mid-training from 10B to 30B (ablation) | OneCAT-1.5B stage-2 · GenEval | GenEval: 80.7 -> 81.2 | Improved |
| increase generation tokens in mid-training (ablation) | OneCAT-1.5B stage-2 · Understanding average | average score: 55.1 -> 56.0 (0B -> 30B), 55.2 at 45B | No clear effect |
| early-fusion decoder-only instead of encoder-based late fusion (no distillation) (ablation) | OneCAT-1.5B stage-1 · 7-benchmark average (70B tokens) | average score: 37.8 -> 38.3 | No clear effect |
| remove vision encoder, use decoder-only multi-scale AR generation vs diffusion BAGEL | OneCAT-3B · 1024x1024 T2I inference | inference time (s): 26.29 -> 2.85 | Improved |
| remove external ViT for understanding prefilling | OneCAT-3B vs Qwen2.5-VL-3B · 1792x1792 input | TTFT (s): 0.583 -> 0.225 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 2 on most shared results.
- OmniGen2 — OmniGen2: Towards Instruction-Aligned Multimodal Generation (CVPR 2026) · better on 0 of 1 shared result
- Show-o2-7B — Show-o: One Single Transformer to Unify Multimodal Understanding and Generation (ICLR 2025) · better on 0 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Janus: Decoupling Visual Encoding for Unified Multimodal Understanding and Generation (CVPR 2025 · similarity 0.91)
- UVU: Improving Multimodal Understanding via Vision-Language Unified Autoregressive Paradigm (CVPR 2026 · similarity 0.88)
- Show-o: One Single Transformer to Unify Multimodal Understanding and Generation (ICLR 2025 · similarity 0.88)
- JanusFlow: Harmonizing Autoregression and Rectified Flow for Unified Multimodal Understanding and Generation (CVPR 2025 · similarity 0.88)
- MANZANO: A Simple and Scalable Unified Multimodal Model with a Hybrid Vision Tokenizer (ICLR 2026 · similarity 0.88)
- OneThinker: All-in-one Reasoning Model for Image and Video (CVPR 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
