Research map / Image generation / Autoregressive and latent generation
Next-scale and token autoregressive images
Generate images autoregressively over scales or continuous tokens with improved tokenizers.
90 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2021–2026) · an approach in Autoregressive and latent generation, keyword cluster “latent space · modeling · vae”
State History Benchmarks Tried Checklist Preprints Who Export
Where it stands
Leveling off Several papers from 2025–26 still report on 1 tracked benchmark, but the best result has barely moved since 2024.
6 papers in 2023–24 and 79 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 1.3% to 8.8%.
What looks settled
- ImageNet 256×256 (FID): best up to 2024 1.73, best in 2025–26 2.21 from 4 papers, at most 0.2% better, if at all.
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
- 2021 · ImageBART: Bidirectional Context with Multinomial Diffusion for Autoregressive Image Synthesis (NeurIPS 2021 · 52 citations)
- 2022 · Autoregressive Image Generation Using Residual Quantization (CVPR 2022 · 265 citations)
- 2023 · Towards Accurate Image Coding: Improved Autoregressive Image Generation With Dynamic Vector Quantization (CVPR 2023 · 26 citations)
- 2024 · Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · 111 citations)
- 2025 · Infinity: Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis (CVPR 2025 · 38 citations)
- 2026 · MVAR: Visual Autoregressive Modeling with Scale and Spatial Markovian Conditioning (ICLR 2026 · 1 citation)
Earliest papers here
- ImageBART: Bidirectional Context with Multinomial Diffusion for Autoregressive Image Synthesis (NeurIPS 2021 · 52 citations)
- BANG: Bridging Autoregressive and Non-autoregressive Generation with Large Scale Pretraining (ICML 2021 · 24 citations)
- Autoregressive Image Generation Using Residual Quantization (CVPR 2022 · 265 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)
5 papers, 2024–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 1.73 | Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · 111 citations) |
What has been tried
32 papers here report 366 measured changes (a component added, removed or swapped, compared with a baseline): helped 66%, no clear effect 7%, hurt 26%, 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 |
|---|---|---|---|---|
| replace AR next-token prediction with VAR next-scale prediction | ImageNet 256×256 · FID | 18.65 -> 5.22 | Helped | Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · 111 citations) |
| Scale the discrete tokenizer vocabulary from Vd=216 to Vd=264 | ImageNet-256 · rFID | 1.22 -> 0.33 | Helped | Infinity: Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis (CVPR 2025 · 38 citations) |
| Replace raster order with random order | ImageNet 256×256, without CFG · FID | 19.23 -> 13.07 | Helped | Autoregressive Image Generation without Vector Quantization (NeurIPS 2024 · 38 citations) |
| Add sequential generation of initial region tokens | ImageNet 256×256 validation · FID | 3.67 -> 2.61 | Helped | Parallelized Autoregressive Visual Generation (CVPR 2025 · 12 citations) |
| Replace shared RoPE-rotated position embedding with dense learnable embeddings | ImageNet 256×256 · FID | 2.82 -> 3.07 | Hurt | RandAR: Decoder-only Autoregressive Visual Generation in Random Orders (CVPR 2025 · 9 citations) |
| Use 6 rather than 8 drafting steps in collaborative decoding | ImageNet1K-256 · speedup | 1.7x -> 2.9x | Helped | Collaborative Decoding Makes Visual Auto-Regressive Modeling Efficient (CVPR 2025 · 7 citations) |
| add DINOv2 feature distillation | Image editing · Text-image alignment | Helped | EditAR: Unified Conditional Generation with Autoregressive Models (CVPR 2025 · 7 citations) | |
| Add MIM loss | ImageNet-256×256, 50 epochs · FID | 31.35 -> 30.58 | Helped | Understand Before You Generate: Self-Guided Training for Autoregressive Image Generation (NeurIPS 2025 · 2 citations) |
| Replace VAR residual prediction with GT prediction using the VAR tokenizer | ImageNet-1K 40-epoch ablation · FID | 4.00 -> N/A | Hurt | FlexVAR: Flexible Visual Autoregressive Modeling without Residual Prediction (NeurIPS 2025 · 1 citation) |
| add frequency-guided residual quantization | ImageNet validation · rFID | 1.62 -> 1.40 | Helped | NFIG: Multi-Scale Autoregressive Image Generation via Frequency Ordering (NeurIPS 2025 · 1 citation) |
| Add residual diffusion to model residual tokens | ImageNet class-conditioned generation, 256px · FID | 2.67 -> 2.39 | Helped | HART: Efficient Visual Generation with Hybrid Autoregressive Transformer (ICLR 2025 · 1 citation) |
| Add a <bkspc> action while training with MLE | OpenWebText · MAUVE | 0.85 ± 0.03 -> 0.84 ± 0.02 | Hurt | SequenceMatch: Imitation Learning for Autoregressive Sequence Modelling with Backtracking (ICLR 2024 · 1 citation) |
What reviewers will expect
Datasets and baselines named in the 29 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-1K · 8 of 29 papers
- ImageNet · 7 of 29 papers
- OpenImages · 4 of 29 papers
- COCOStuff · 2 of 29 papers
- MultiGen-20M · 2 of 29 papers
- LAION · 2 of 29 papers
- MS-COCO · 2 of 29 papers
- ADE20K · 2 of 29 papers
Baselines
- VAR · 11 of 29 papers
- DiT-XL/2 · 8 of 29 papers
- MaskGIT · 8 of 29 papers
- LlamaGen · 7 of 29 papers
- MAR · 6 of 29 papers
- VQGAN · 5 of 29 papers
- RQTransformer · 4 of 29 papers
- StyleGAN-XL · 4 of 29 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 85 papers; from OpenAlex and the papers' affiliations).
- ByteDance · 5 papers
- The University of Hong Kong · 5 papers
- Shanghai Jiao Tong University · 4 papers
- University of Science and Technology of China · 3 papers
- Peking University · 3 papers
- Shanghai AI Laboratory · 3 papers
- Institute of Artificial Intelligence · 2 papers
- Massachusetts Institute of Technology · 2 papers
- Google · 2 papers
- Johns Hopkins University · 2 papers
Use it
Download BibTeX of 27 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)
- Accelerating autoregressive image generation (22 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).