Research map / Image generation / Autoregressive and latent generation
DPAR: Dynamic Patchification for Efficient Autoregressive Visual Generation
Divyansh Srivastava, Akshay Mehra, Pranav Maneriker, Debopam Sanyal, Vishnu Raj, Vijay Kamarshi, Fan Du, Joshua Kimball
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 |
|---|---|---|---|
| add entropy gating alone to static patches (ablation) | DPAR-L · ImageNet 256 (DPAR-L, 50 epochs) | FID-50K: 3.58 -> 3.91 | Hurt |
| add max patch length on top of entropy gating (ablation) | DPAR-L · ImageNet 256 (DPAR-L, 50 epochs) | FID-50K: 3.91 -> 3.45 | Improved |
| add row-boundary resets (ablation) | DPAR-L · ImageNet 256 (DPAR-L, 50 epochs) | FID-50K: 3.45 -> 3.32 | Improved |
| full entropy patchification vs static patches (ablation) | DPAR-L · ImageNet 256 (DPAR-L, 50 epochs) | FID-50K: 3.58 -> 3.32 | Improved |
| use entropy threshold 7.8 instead of 8.0 (ablation) | DPAR-L · ImageNet 256 (DPAR-L, 50 epochs) | FID-50K: 3.43 -> 3.32 | Mixed |
| raise max patch length from 1 to 4 (ablation) | DPAR-L · ImageNet 256 (DPAR-L, 50 epochs) | FID-50K: 3.73 -> 3.32 | Improved |
| raise max patch length from 4 to 16 (ablation) | DPAR-L · ImageNet 256 (DPAR-L, 50 epochs) | FID-50K: 3.32 -> 3.61 | Hurt |
| raise inference entropy threshold to 8.1 (dynamic training) (ablation) | DPAR-L · ImageNet 256 (DPAR-L, 50 epochs) | FID-50K: 3.32 -> 3.39 | No clear effect |
| raise inference patch length to 2.16 (static training) (ablation) | Static patch model · ImageNet 256 (DPAR-L, 50 epochs) | FID-50K: 3.58 -> 25.59 | Hurt |
| train with dynamic patches rather than raster tokens (linear probe) | DPAR-L vs LlamaGen-L · ImageNet linear probe | Top-1 accuracy (%): 32.62 -> 37.82 | Improved |
| add repeated start/end encodings to dynamic RoPE (ablation) | DPAR-L · ImageNet 256 (DPAR-L, 50 epochs) | FID-50K: 3.42 -> 3.31 | Improved |
| use 1 encoder layer and 4 decoder layers instead of 4 and 1 (ablation) | DPAR-L · ImageNet 256 (DPAR-L, 50 epochs) | FID-50K: 3.85 -> 3.32 | Improved |
Closest papers by meaning
- Vision Foundation Models as Effective Visual Tokenizers for Autoregressive Generation (NeurIPS 2025 · similarity 0.90)
- Parallelized Autoregressive Visual Generation (CVPR 2025 · similarity 0.90)
- Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots (ICML 2025 · similarity 0.90)
- RandAR: Decoder-only Autoregressive Visual Generation in Random Orders (CVPR 2025 · similarity 0.90)
- ZipAR: Parallel Autoregressive Image Generation through Spatial Locality (ICML 2025 · similarity 0.90)
- BitDance: Scaling Autoregressive Generative Models with Binary Tokens (NeurIPS 2026 · 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.
