Research map / Image generation / Text-to-image diffusion
Progress by Pieces: Test-Time Scaling for Autoregressive Image Generation
Joonhyung Park, Hyeongwon Jang, Joowon Kim, Eunho Yang
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 Best-of-N with GridAR grid-partitioned progressive generation (N=4 vs N=8) | Janus-Pro-7B · T2I-CompBench++ | average score: 14.4% higher than Best-of-N (N=8) with 25.6% less wall-clock time | Improved |
| apply GridAR to LlamaGen at same N | LlamaGen · T2I-CompBench++ | average score: 4.9% | Improved |
| replace Best-of-N (N=8) with GridAR (N=8) on LlamaGen color attribution | LlamaGen · GenEval Color Attribution | score: 0.14 -> 0.13 | No clear effect |
| add GridAR to EditAR instead of Best-of-N (N=4) | EditAR · PIE-Bench | structure distance and MSE: -10.54% structure distance, -26.07% MSE | Improved |
| GridAR (N=4) vs Best-of-N (N=8) on instruction fidelity in editing | EditAR · PIE-Bench | CLIP similarity: 25.532 vs 25.628 | No clear effect |
| add mid-generation layout-specified prompt reformulation (ablation) | GridAR (N=4) Janus-Pro · T2I-CompBench++ | average score: 0.4764 -> 0.5654 | Improved |
| reformulate prompt at the start with a planner instead of midway from partial images (ablation) | GridAR (N=4) Janus-Pro · T2I-CompBench++ | average score: 0.5654 -> 0.5027 | Hurt |
| swap GPT-4.1 verifier for GPT-4o-mini (ablation) | GridAR (N=4) · T2I-CompBench++ | improvement over Best-of-N: 18.5% improvement | No clear effect |
| swap GPT-4.1 verifier for open-source MiniCPM-V 4.5 (ablation) | GridAR (N=4) · T2I-CompBench++ | improvement over Best-of-N: 15.8% | No clear effect |
| use verification step in GridAR rather than plain Best-of-N sampling | GridAR (N=4) · Janus-Pro wall-clock time | seconds: 23.9s -> 36.5s (vs Best-of-N N=4) | Hurt |
| reject only infeasible candidates instead of top-k selection (ablation) | GridAR · candidate pool diversity | design rationale | Improved |
Closest papers by meaning
- Hawk: Leveraging Spatial Context for Faster Autoregressive Text-to-Image Generation (NeurIPS 2025 · similarity 0.91)
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.90)
- Fluid: Scaling Autoregressive Text-to-image Generative Models with Continuous Tokens (ICLR 2025 · similarity 0.90)
- RandAR: Decoder-only Autoregressive Visual Generation in Random Orders (CVPR 2025 · similarity 0.90)
- TTS-VAR: A Test-Time Scaling Framework for Visual Auto-Regressive Generation (NeurIPS 2025 · similarity 0.90)
- LANTERN: Accelerating Visual Autoregressive Models with Relaxed Speculative Decoding (ICLR 2025 · similarity 0.90)
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.
