Research map / Image generation / Autoregressive and latent generation
Autoregression with Self-Token Prediction
Dengsheng Chen, Yangming Shi, Enhua Wu
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 non-informative shared embedding with neighborhood-averaging input h(.) in Manhattan grouping (ablation) | SAR-L · ImageNet-1K 256x256 (SAR-L, 250K iters) | FID-50K: 58.49 -> 4.40 | Improved |
| switch from next-token to next-row grouping (ablation) | SAR-L · ImageNet-1K 256x256 (SAR-L, 250K iters) | FID-50K: 4.92 -> 22.78 | Hurt |
| switch from next-token to next-layer grouping (ablation) | SAR-L · ImageNet-1K 256x256 (SAR-L, 250K iters) | FID-50K: 4.92 -> 29.33 | Hurt |
| switch from next-token to Manhattan neighbor grouping with AggAVG (ablation) | SAR-L · ImageNet-1K 256x256 (SAR-L, 250K iters) | Latency: 622ms -> 106ms | Improved |
| switch from next-token to Manhattan neighbor grouping with AggAVG (ablation) | SAR-L · ImageNet-1K 256x256 (SAR-L, 250K iters) | FID-50K: 4.92 -> 4.40 | Improved |
| add classifier-free guidance (scale 3.50) (ablation) | SAR-L · ImageNet-1K 256x256 (SAR-L, guidance ablation) | FID-50K: 39.68 -> 4.40 | Improved |
| add x-CFG alone (ablation) | SAR-L · ImageNet-1K 256x256 (SAR-L, guidance ablation) | FID-50K: 39.68 -> 18.53 | Improved |
| replace standard CFG with combined x-CFG and z-CFG (ablation) | SAR-L · ImageNet-1K 256x256 (SAR-L, guidance ablation) | FID-50K: 4.40 -> 3.56 | Improved |
| replace standard CFG with combined x-CFG and z-CFG (ablation) | SAR-L · ImageNet-1K 256x256 (SAR-L, guidance ablation) | IS: 307.34 -> 275.13 | Hurt |
| use flow matching head instead of DDPM objective (ablation) | SAR-L · ImageNet-1K 256x256 (SAR-L, denoising space) | FID-50K: 29.48 -> 3.56 | Improved |
| use KL-VAE latent instead of raw RGB pixel tokens (ablation) | SAR-L · ImageNet-1K 256x256 (SAR-L, denoising space) | FID-50K: 8.36 -> 3.56 | Improved |
| use KL-VAE latent instead of joint image-video VAE latent (ablation) | SAR-L · ImageNet-1K 256x256 (SAR-L, denoising space) | FID-50K: 4.84 -> 3.56 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- MaskGIT — MaskGIT: Masked Generative Image Transformer (CVPR 2022) · better on 1 of 1 shared result
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
- Parallelized Autoregressive Visual Generation (CVPR 2025 · similarity 0.90)
- Understand Before You Generate: Self-Guided Training for Autoregressive Image Generation (NeurIPS 2025 · similarity 0.90)
- Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots (ICML 2025 · similarity 0.90)
- Autoregressive Image Generation without Vector Quantization (NeurIPS 2024 · similarity 0.88)
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.88)
- ControlAR: Controllable Image Generation with Autoregressive Models (ICLR 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.
