Research map / Image generation / Autoregressive and latent generation
Patchwise Generative ConvNet: Training Energy-Based Models From a Single Natural Image for Internal Learning
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use PatchGenCN instead of SinGAN for single-image generation | Patchwise image generation on Places · Places | SIFID: 0.11 -> 0.09 | Helped |
| Use PatchGenCN instead of SinGAN for single-image generation | Patchwise image generation on Places · Places | paired real-vs-fake samples labeled real: 35.73% -> 33.60% | Hurt |
| Use PatchGenCN instead of SinGAN for single-image generation | Patchwise image generation on Places · Places | unpaired real-vs-fake samples labeled real: 39.33% -> 40.13% | Helped |
| Use PatchGenCN instead of SinGAN for single-image generation | Patchwise image generation on Places · Places | NIQE: 5.22 -> 5.10 | Helped |
| Use PatchGenCN instead of the listed baselines | 4× single-image super-resolution · BSD100, example 1 | NIQE: 5.59 -> 4.32; 6.83 -> 4.32; 4.40 -> 4.32; 5.21 -> 4.32 | Helped |
| Use PatchGenCN instead of the listed baselines | 4× single-image super-resolution · BSD100, example 2 | NIQE: 6.49 -> 3.74; 4.48 -> 3.74; 4.75 -> 3.74; 4.35 -> 3.74 | Helped |
| Use PatchGenCN instead of the listed baselines | 4× single-image super-resolution · BSD100, example 1 | PSNR: 25.80, 28.79, 28.35, 24.36 -> 22.54 | Mixed |
| Use PatchGenCN instead of the listed baselines | 4× single-image super-resolution · BSD100, example 2 | PSNR: 31.86, 31.72, 33.54, 30.56 -> 28.45 | Mixed |
| Add higher-resolution scales to the model pyramid (ablation) | PatchGenCN image generation · Single-image synthesis | visual detail and structure | Helped |
| Add a self-supervised task during training (λ = 0.1 rather than λ = 0) (ablation) | PatchGenCN image generation · Single-image synthesis | synthesis quality | Helped |
| Set the self-supervised task weight either too large or too small (ablation) | PatchGenCN training · Single-image synthesis | training stability | Hurt |
Closest papers by meaning
- A Theory of Generative ConvNet (ICML 2016 · similarity 0.90)
- Learning Generative ConvNets via Multi-Grid Modeling and Sampling (CVPR 2018 · similarity 0.90)
- Efficient and Training-Free Single-Image Diffusion Models (CVPR 2026 · similarity 0.88)
- Learning Energy-Based Generative Models via Coarse-to-Fine Expanding and Sampling (ICLR 2021 · similarity 0.87)
- Meta Internal Learning (NeurIPS 2021 · similarity 0.87)
- Patched Denoising Diffusion Models For High-Resolution Image Synthesis (ICLR 2024 · similarity 0.86)
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.