Research map / Image generation / Autoregressive and latent generation
(1D) Ordered Tokens Enable Efficient Test-Time Search
Zhitong Gao, Parham Rezaei, Ali Cy, Mingqiao Ye, Nataša Jovanović, Jesse Allardice, Afshin Dehghan, Amir Zamir, Roman Bachmann, Oğuzhan Fatih Kar
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 |
|---|---|---|---|
| use beam search on 1D ordered tokens (FlexTok) versus 2D grid tokens (ablation) | AR text-to-image models with matched data, architecture and compute · COCO Karpathy val (300-image subset) | CLIPScore vs inference compute | Improved |
| use best-of-N and lookahead search on 1D ordered vs 2D grid tokens (ablation) | AR text-to-image models · COCO Karpathy val (300-image subset) | CLIPScore vs inference compute | No clear effect |
| add beam search with DreamSim verifier over first 32 tokens | FlexTok · DreamBench++ | DINO-I: 32.5 -> 50.9 | Improved |
| add lookahead search with DreamSim verifier | Janus-1.3B · DreamBench++ | DINO-I: 34.8 -> 40.7 | Improved |
| add beam search with DreamSim verifier | FlexTok · DreamBench++ | CLIP-T (prompt following): 34.1 -> 33.1 | No clear effect |
| add search on top of conditional prior (ablation) | FlexTok · GenEval subset, two object | accuracy (%): 48 -> 81 | Improved |
| replace conditional AR prior with uniform prior in search (ablation) | FlexTok · GenEval subset, two object | accuracy (%): 81 -> 32 | Hurt |
| replace uniform prior with unconditional AR prior in search (ablation) | FlexTok · GenEval subset, single object | accuracy (%): 79 -> 85 | Improved |
| scale AR model size with best-of-N under limited vs large inference compute (ablation) | FlexTok AR (212M to 3.4B) · GenEval | overall accuracy | Mixed |
| use an ensemble of eight verifiers instead of a single verifier (ablation) | FlexTok beam search · GenEval | average rank across metrics: 2.12 average rank | Improved |
| use any verifier in beam search versus no search (ablation) | FlexTok · GenEval | evaluation metrics | Improved |
Closest papers by meaning
- Fluid: Scaling Autoregressive Text-to-image Generative Models with Continuous Tokens (ICLR 2025 · similarity 0.88)
- Vision Foundation Models as Effective Visual Tokenizers for Autoregressive Generation (NeurIPS 2025 · similarity 0.88)
- Parallelized Autoregressive Visual Generation (CVPR 2025 · similarity 0.87)
- ImageFolder: Autoregressive Image Generation with Folded Tokens (ICLR 2025 · similarity 0.87)
- When Worse is Better: Navigating the Compression Generation Trade-off In Visual Tokenization (NeurIPS 2025 · similarity 0.87)
- MARS: Enabling Autoregressive Models Multi-Token Generation (NeurIPS 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.
