Research map / Image generation / Autoregressive and latent generation
Semantic Context Matters: Improving Conditioning for Autoregressive Models
Dongyang Jin, Ryan Xu, Jianhao Zeng, Rui Lan, Yancheng Bai, Lei Sun, Xiangxiang Chu
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 |
|---|---|---|---|
| replace plain resize with learnable semantic compression plus L_pres (4x) (ablation) | SCAR · MultiGen-20M Depth | FID: 15.78 -> 14.70 | Improved |
| add semantic preservation loss L_pres to learnable compression (ablation) | SCAR · MultiGen-20M HED | FID: 9.89 -> 9.43 | Improved |
| learnable compression versus PixelUnshuffle (ablation) | SCAR · MultiGen-20M HED | FID: 9.82 vs 9.89 | No clear effect |
| 4x compression instead of no compression (ablation) | SCAR · MultiGen-20M HED | FID: 9.29 -> 9.43 | No clear effect |
| 16x compression instead of 4x (ablation) | SCAR · MultiGen-20M Depth | FID: 14.70 -> 16.10 | Hurt |
| use DINOv2-B instead of CLIP-B encoder (ablation) | SCAR · MultiGen-20M HED | FID: 13.78 -> 9.43 | Improved |
| scale image encoder from DINOv2-S to DINOv2-B (ablation) | SCAR · MultiGen-20M Depth | FID: 15.26 -> 14.70 | Improved |
| use DINOv2-B instead of ViT-B encoder (ablation) | SCAR · MultiGen-20M HED | SSIM: 77.45 -> 81.76 | Improved |
| add Semantic Alignment Guidance L_align (ablation) | SCAR · instruction editing (qualitative) | instruction fidelity | Improved |
| raise alignment weight delta to 1.0 (ablation) | SCAR · instruction editing (qualitative) | visual quality | Hurt |
| use prefilling (SCAR) rather than decoding-stage injection (ControlAR*) | LlamaGen-XL · PIE-Bench | Structure distance: 116.99 -> 30.98 | Improved |
| SCAR versus EditAR on same backbone | LlamaGen-XL · PIE-Bench | MSE: 130.27 -> 83.47 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- ControlAR* — ControlAR: Controllable Image Generation with Autoregressive Models (ICLR 2025) · better on 8 of 8 shared results
- EditAR — EditAR: Unified Conditional Generation with Autoregressive Models (CVPR 2025) · better on 4 of 4 shared results
- ControlNet++ — ControlNet++: Improving Conditional Controls with Efficient Consistency Feedback (ECCV 2024) · better on 3 of 3 shared results
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
- Context-Aware Autoregressive Image Generation for Emerging Reasoning Properties (NeurIPS 2026 · similarity 0.90)
- EditAR: Unified Conditional Generation with Autoregressive Models (CVPR 2025 · similarity 0.89)
- Understand Before You Generate: Self-Guided Training for Autoregressive Image Generation (NeurIPS 2025 · similarity 0.89)
- Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots (ICML 2025 · similarity 0.88)
- Toward Guidance-Free AR Visual Generation via Condition Contrastive Alignment (ICLR 2025 · similarity 0.88)
- Vision Foundation Models as Effective Visual Tokenizers for Autoregressive Generation (NeurIPS 2025 · similarity 0.88)
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.
