Research map / Image generation / Autoregressive and latent generation
ZipAR: Parallel Autoregressive Image Generation through Spatial Locality
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ZipAR, a training-free parallel decoding framework that lets autoregressive image generators decode spatially adjacent tokens across rows, using adaptive local windows with speculative-style rejection sampling, cutting forward passes by up to 91% on Emu3-Gen.
Problem
Autoregressive visual generation decodes tokens one per forward pass in raster order, making high-resolution image generation slow; existing LLM parallel decoding methods ignore visual spatial structure.
Approach
Exploits spatial locality of image tokens: once a row has decoded enough tokens beyond a local window, the next row starts decoding in parallel; window size is adaptively assigned per token via a rejection-sampling acceptance criterion analogous to speculative decoding, with no retraining.
Key results
- On Emu3-Gen, ZipAR reduces forward steps by up to 91%, with 75% reduction on Lumina-mGPT-7B and 81% on LlamaGen-XL (MS-COCO setting) with minimal quality loss.
- On ImageNet with LlamaGen-XL, ZipAR-12 reaches FID 3.67 in 331 steps versus SJD's 3.87 in 335 steps.
- For LlamaGen-XL-512 on MS-COCO, ZipAR-15 cuts steps by 45.1% with CLIP score unchanged at 0.287.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply ZipAR with minimum window size 16 | LlamaGen-L · ImageNet 256×256 | forward steps: 576 -> 422 (-26.7%) | Helped |
| Apply ZipAR with minimum window size 16 | LlamaGen-L · ImageNet 256×256 | FID: 3.16 -> 3.14 | Helped |
| Apply ZipAR with minimum window size 12 | LlamaGen-XL · ImageNet 256×256 | forward steps: 576 -> 331 (-41.8%) | Helped |
| Apply ZipAR with minimum window size 12 | LlamaGen-XL · ImageNet 256×256 | FID: 2.83 -> 3.67 | Hurt |
| Apply ZipAR-15 | LlamaGen-XL-512 · MS-COCO | CLIP Score: 0.287 -> 0.287 | No clear effect |
| Apply ZipAR-15 | LlamaGen-XL-512 · text-to-image | VQAScore: 0.6439 -> 0.6534 | Helped |
| Apply ZipAR-20 | Lumina-mGPT-768 · text-to-image | VQAScore: 0.6579 -> 0.6595 | Helped |
| Apply ZipAR-3 | LlamaGen-XL-512 · MS-COCO | forward steps: 1024 -> 185 (-81.9%) | Helped |
| Reduce the minimum window size from 15 to 3 | LlamaGen-XL-512 ZipAR · MS-COCO | CLIP Score: 0.287 -> 0.281 | Hurt |
| Use adaptive rather than fixed window size (ablation) | ZipAR on LlamaGen-L · ImageNet | FID | Helped |
| Use classifier-free guidance 1.5 instead of 2.0 (ablation) | ZipAR-16 on LlamaGen-L · ImageNet 256×256 | FID: 3.14 -> 6.18 | Hurt |
| Use classifier-free guidance 1.75 instead of 2.0 (ablation) | ZipAR-16 on LlamaGen-L · ImageNet 256×256 | FID: 3.14 -> 3.72 | Hurt |
Surprising findings
- Visual token attention concentrates on tokens in the same column of the previous row, so raster-order dependency is suboptimal.
- Acceleration ratio is higher for text-to-image models at larger resolution than for class-conditional LlamaGen-L.
Limitations
- Smaller window sizes degrade image quality (e.g., ZipAR-3 lowers CLIP and VQAScore), trading off speed and fidelity.
- Requires model-specific handling of the end-of-row token for models lacking dynamic-resolution end tokens.
- Gains depend on window size tuning; reported results are with hyperparameter choices per model.
Experimental setup
Datasets: ImageNet COCO Parti
Benchmarks: ImageNet 256x256 MS-COCO VQAScore HPSv2 ImageReward Aesthetic Score
Metrics: FID CLIP Score forward steps latency VQAScore HPSv2 ImageReward Aesthetic Score
Baselines: NTP (next-token prediction) SJD Medusa Jacobi decoding
Builds on: speculative decoding Jacobi decoding Medusa Lumina-mGPT LlamaGen
Closest papers by meaning
- RandAR: Decoder-only Autoregressive Visual Generation in Random Orders (CVPR 2025 · similarity 0.92)
- Hawk: Leveraging Spatial Context for Faster Autoregressive Text-to-Image Generation (NeurIPS 2025 · similarity 0.90)
- Parallelized Autoregressive Visual Generation (CVPR 2025 · similarity 0.89)
- Towards Better & Faster Autoregressive Image Generation: From the Perspective of Entropy (NeurIPS 2025 · similarity 0.89)
- LANTERN: Accelerating Visual Autoregressive Models with Relaxed Speculative Decoding (ICLR 2025 · similarity 0.89)
- Collaborative Decoding Makes Visual Auto-Regressive Modeling Efficient (CVPR 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.