PRA Enhances Pixel-Space Autoregressive Image Generation
机器之心 · wechat · 2026-07-13
### PRA: Improving Pixel-Space Autoregressive Image Generation Peking University and DP Technology proposed **Parallel Rollout Approximation (PRA)** to address the suboptimal long-term performance of pixel-space autoregressive (AR) image generation. Systematic experiments reveal two primary bottlenecks in pixel-space AR: - **Difficulty in single-step prediction of high-dimensional pixel tokens**: Higher token dimensions lead to more pronounced FID degradation. - **Training-inference inconsistency**: Training relies on clean prefixes, while inference depends on self-generated "imperfect inputs," causing error accumulation. ### Core Approach PRA revamps both ends of the process: - **Output side**: Instead of directly predicting high-dimensional pixel patches, it first predicts a **low-dimensional intermediate state** before decoding back to pixel space, reducing single-step generation difficulty. - **Input side**: During training, it uses parallel-constructed decoded pixel inputs to approximate the input distribution during rollout, mitigating the teacher-forcing discrepancy. ### Experimental Results On the ImageNet-1K 256×256 class-conditional generation task, PRA significantly boosted pixel-space AR performance: - **PRA-Small (135M)**: Achieved an FID of **2.58**, outperforming the 1.9B parameter FARMER-1.9B/8 (FID 3.60). - **PRA-Base (250M)**: Achieved an FID of **2.21**. - **PRA-Large (511M)**: Achieved an FID of **1.94**, setting a new state-of-the-art for pixel-space AR. ### Additional Findings The paper also evaluates representation capabilities on ImageNet linear probing, where **PRA-Large achieved a top-1 accuracy of 68.80%**. This demonstrates that end-to-end pixel-space AR can learn robust visual representations, not just generate images.
More from Multimodal
- Creator makes a dark-fantasy short film teaser with Google Flow visuals — AI_Cyborg · 2026-07-21
- HarmoHOI generates multi-view hand-object videos and aligned 3D motion in one diffusion model — cn-scut · 2026-07-21
- Open-source Gradio app merges Krea 2 Turbo LoRAs on 6GB systems — Fluid_Kaleidoscope17 · 2026-07-21
- Early Krea2 Gradio WebUI targets 6GB low-VRAM local runs — Fluid_Kaleidoscope17 · 2026-07-21
- Open-source B-roll skill turns scripts into 5-second vertical clips with Codex and Gemini — yangyi · 2026-07-21
- Midjourney 8.2 preview shows explosive glitch-style dog portraits — michaelrabone · 2026-07-21