Research map / Image generation / Autoregressive and latent generation
Accelerating autoregressive image generation
Speed autoregressive and masked image generators with speculative decoding and hierarchical modeling.
22 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2023–2026) · an approach in Autoregressive and latent generation, keyword cluster “latent space · modeling · vae”
State History Benchmarks Tried Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 1 of 1 tracked benchmark.
2 papers in 2023–24 and 20 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 0.4% to 2.2%.
Open questions the data points to
- ImageNet 256×256 (FID): best went from 3.22 (up to 2024) to 1.57 in 2025 — Effective and Efficient Masked Image Generation Models (ICML 2025)
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2023 · Long Horizon Temperature Scaling (ICML 2023 · 3 citations)
- 2024 · Unlocking the Capabilities of Masked Generative Models for Image Synthesis via Self-Guidance (NeurIPS 2024 · 1 citation)
- 2025 · Towards Better & Faster Autoregressive Image Generation: From the Perspective of Entropy (NeurIPS 2025 · 2 citations)
- 2026 · CSD: Content-aware Speculative Decoding for Efficient Image Generation (ICML 2026 · 0 citations)
Earliest papers here
- Long Horizon Temperature Scaling (ICML 2023 · 3 citations)
- Unlocking the Capabilities of Masked Generative Models for Image Synthesis via Self-Guidance (NeurIPS 2024 · 1 citation)
- Towards Better & Faster Autoregressive Image Generation: From the Perspective of Entropy (NeurIPS 2025 · 2 citations)
Benchmark records
The best number each year that beat every earlier one, among papers of this approach reporting their own method. Numbers are extracted by AI from each paper's text; check the paper before citing.
ImageNet 256×256 · FID (lower is better)
4 papers, 2024–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 3.22 | Unlocking the Capabilities of Masked Generative Models for Image Synthesis via Self-Guidance (NeurIPS 2024 · 1 citation) |
| 2025 | 1.57 | Effective and Efficient Masked Image Generation Models (ICML 2025) |
What has been tried
13 papers here report 157 measured changes (a component added, removed or swapped, compared with a baseline): helped 71%, no clear effect 9%, hurt 20%, mixed 1%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| use Markovian next-scale conditioning | ImageNet 256×256 · FID | 3.50 -> 3.76 | Hurt | HMAR: Efficient Hierarchical Masked Auto-Regressive Image Generation (CVPR 2025 · 1 citation) |
| predict cumulative embeddings instead of discrete tokens directly in parallel across depths | ImageNet 256×256 · FID (w/ CFG) | 2.91 -> 2.43 | Helped | Efficient Generative Modeling with Residual Vector Quantization-Based Tokens (ICML 2025 · 1 citation) |
| Increase re-masking proportion from 0.1 to 0.7 | ARC-C · accuracy | 45.9 -> 47.8 | Helped | Adaptive Classifier-Free Guidance via Dynamic Low-Confidence Masking (NeurIPS 2025 · 1 citation) |
| Use Gaussian blur in logit space for guidance | ImageNet 256×256 · FID | 8.32 -> 3.22 | Hurt | Unlocking the Capabilities of Masked Generative Models for Image Synthesis via Self-Guidance (NeurIPS 2024 · 1 citation) |
| Use adaptive rather than fixed window size | ImageNet · FID | Helped | ZipAR: Parallel Autoregressive Image Generation through Spatial Locality (ICML 2025 · 0 citations) | |
| Use NGS during the last 2K evaluations instead of spending all 10K on training | Molecular benchmark average · Top-10 score | 0.826 -> 0.835 | Helped | Neural Genetic Search in Discrete Spaces (ICML 2025 · 0 citations) |
| Use 8 Local-FEs instead of Full-FEs (u=0) | ImageNet256 class-conditional, no CFG, 16 steps · FID | 4.46 -> 5.02 | Hurt | Plug-and-Play Context Feature Reuse for Efficient Masked Generation (NeurIPS 2025 · 0 citations) |
| Add embedding normalization to the denoising process | Image generation · Visual coherence | Helped | Speculative Jacobi-Denoising Decoding for Accelerating Autoregressive Text-to-image Generation (NeurIPS 2025 · 0 citations) | |
| Use latent-neighbor selection instead of random selection | MS-COCO validation, sampling (τ=1) · mean accepted length | 1.26 -> 2.09 | Helped | LANTERN: Accelerating Visual Autoregressive Models with Relaxed Speculative Decoding (ICLR 2025 · 0 citations) |
| Increase the number of training iterations | Few-step generation · FID | Helped | Distilled Decoding 1: One-step Sampling of Image Auto-regressive Models with Flow Matching (ICLR 2025 · 0 citations) | |
| Use the exp mask schedule instead of the linear schedule | ImageNet 256×256, epoch 400 · FID | 24.96 -> 11.90 | Helped | Effective and Efficient Masked Image Generation Models (ICML 2025) |
What reviewers will expect
Datasets and baselines named in the 14 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- ImageNet · 5 of 14 papers
- MS-COCO 2014 · 2 of 14 papers
- ImageNet256 · 2 of 14 papers
- MSCOCO 2017 · 2 of 14 papers
- LAION-COCO · 2 of 14 papers
Baselines
- MaskGIT · 5 of 14 papers
- DiT-XL/2 · 3 of 14 papers
- SJD · 3 of 14 papers
- VAR · 3 of 14 papers
- MAR · 3 of 14 papers
- MAGE · 2 of 14 papers
- Medusa · 2 of 14 papers
- Jacobi decoding · 2 of 14 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 22 papers; from OpenAlex and the papers' affiliations).
- NVIDIA · 3 papers
- The Chinese University of Hong Kong · 3 papers
- Alibaba Group · 3 papers
- Stanford University · 2 papers
- Zhejiang University · 2 papers
- Korea Advanced Institute of Science and Technology · 2 papers
- Peking University · 2 papers
Use it
Download BibTeX of 21 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Generative latent space tooling (129 papers)
- High-resolution diffusion synthesis (140 papers)
- Latent diffusion with learned tokenizers (114 papers)
- Hierarchical VQ-VAE generation (62 papers)
- Next-scale and token autoregressive images (90 papers)
Elsewhere in the topic
- Classical generative image synthesis (147 papers)
- Retrieval and matching with generative features (140 papers)
- Generalizable fake image detection (119 papers)
- Synthetic training data from generators (111 papers)
- Compositional scene generation (105 papers)
- Robust diffusion watermarking (79 papers)
- Neural text-to-speech synthesis (72 papers)
- Equivariant diffusion for 3D molecules (71 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).