Research map / Image generation / Autoregressive and latent generation
Parallelized Autoregressive Visual Generation
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 PAR, which generates spatially distant tokens with weak dependencies in parallel while keeping sequential generation for initial and local tokens, speeding up autoregressive image and video generation by 3.6-9.5x on ImageNet and UCF-101.
Problem
Autoregressive visual generation is slow because it predicts tokens one at a time, while naive parallel prediction of adjacent visual tokens with strong dependencies produces inconsistent, distorted local structures.
Approach
Group tokens at corresponding positions across spatially distant regions for simultaneous prediction, generate each region's initial tokens sequentially to set global structure, and use group-wise bidirectional attention with causal attention between groups, all within a standard AR transformer with learnable transition tokens and 2D RoPE.
Key results
- PAR-3B-4x reaches FID 2.29 in 147 steps and 3.46s per image vs. LlamaGen-3B FID 2.18 at 576 steps and 12.41s (about 3.6x speedup).
- PAR-3B-16x reaches FID 2.88 in 51 steps and 1.31s per image (about 9.5x speedup).
- On UCF-101, PAR-16x achieves FVD 103.4 with 3.44s vs. 43.30s for the token-by-token PAR-1x baseline (FVD 94.1), about 12.6x speedup.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add sequential generation of initial region tokens (ablation) | PAR-XL · ImageNet 256×256 validation | FID: 3.67 -> 2.61 | Helped |
| Increase parallel group size from n=1 to n=4 (ablation) | PAR-XXL · ImageNet 256×256 validation | FID: 2.34 -> 2.35 | No clear effect |
| Increase parallel group size from n=1 to n=16 (ablation) | PAR-XXL · ImageNet 256×256 validation | FID: 2.34 -> 3.02 | Hurt |
| Use full attention within parallel groups instead of causal attention (ablation) | PAR-XL · ImageNet 256×256 validation | FID: 3.64 -> 2.61 | Helped |
| Use distant rather than raster ordering for single-token prediction (ablation) | PAR-XL · ImageNet 256×256 validation | FID: 2.62 -> 2.64 | No clear effect |
| Use distant rather than raster ordering for multi-token prediction (ablation) | PAR-XL · ImageNet 256×256 validation | FID: 5.64 -> 2.61 | Helped |
| Increase model size from 343M to 3.1B parameters (ablation) | PAR-4× · ImageNet 256×256 validation | FID: 3.76 -> 2.29 | Helped |
| Generate 4 tokens in parallel instead of token-by-token (ablation) | PAR video generation · UCF-101 | FVD: 94.1 -> 99.5 | Hurt |
| Generate 16 tokens in parallel instead of token-by-token (ablation) | PAR video generation · UCF-101 | FVD: 94.1 -> 103.4 | Hurt |
Surprising findings
- Naive raster-order multi-token prediction degrades FID to 5.64, while distant-region ordering keeps FID at 2.61 with the same 147 steps.
- Parallel decoding of initial tokens across regions without sequential initialization causes repeated or conflicting global structures.
Limitations
- Parallel generation along the temporal dimension for video was less effective and left to future work.
- Quality degrades somewhat with more aggressive parallelization (FID rises by 0.67 at n=16 for PAR-XXL), and smaller models show a larger gap to the sequential baseline.
Experimental setup
Datasets: ImageNet-1K UCF-101
Benchmarks: ImageNet 256x256 UCF-101
Metrics: FID IS Precision Recall FVD generation steps wall-clock time
Baselines: LlamaGen MaskGIT VAR MAR MAGVIT-v2-AR OmniTokenizer DiT-XL/2 ADM
Builds on: LlamaGen VQGAN MAGVIT-v2 RoPE
Closest papers by meaning
- Vision Foundation Models as Effective Visual Tokenizers for Autoregressive Generation (NeurIPS 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)
- VideoMAR: Autoregressive Video Generation with Continuous Tokens (NeurIPS 2025 · similarity 0.89)
- ZipAR: Parallel Autoregressive Image Generation through Spatial Locality (ICML 2025 · similarity 0.89)
- OmniGen-AR: AutoRegressive Any-to-Image Generation (NeurIPS 2025 · similarity 0.89)
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.