Research map / Image generation / Autoregressive and latent generation
Learning to See by Looking at Noise
Manel Baradad, Jonas Wulff, Tongzhou Wang, Phillip Isola, Antonio Torralba
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 |
|---|---|---|---|
| Pretrain with StyleGAN-Oriented synthetic noise rather than random initialization | ResNet-50 · ImageNet-1K | top-1 accuracy: 4.36 -> 38.12 | Improved |
| Combine Dead leaves - Shapes, Feature Vis. - Dead leaves, Spectrum + Color + WMM, and StyleGAN-Oriented | ResNet-50 · ImageNet-1K | top-1 accuracy: 38.12 -> 40.06 | Improved |
| Continuously sample StyleGAN-Oriented images instead of training on a fixed 1.3M images | ResNet-50 with StyleGAN-Oriented · ImageNet-1K | top-1 accuracy: 38.12% -> 38.94% | Improved |
| Add diverse synthetic procedural-noise data rather than use random CNN features | ResNet-50 · ImageNet-100 | top-1 accuracy: 10.84 -> 60.66 | Improved |
| Use higher-frequency noise maps and wavelet filters rather than default StyleGAN initialization (ablation) | AlexNet · ImageNet-100 | top-1 accuracy: 0.34 -> 0.42 | Improved |
| Add sparse noise envelope and random convolution biases rather than default StyleGAN initialization (ablation) | AlexNet · ImageNet-100 | top-1 accuracy: 0.34 -> 0.40 | Improved |
| Tie wavelet filters across output channels rather than use default StyleGAN initialization (ablation) | AlexNet · ImageNet-100 | top-1 accuracy: 0.34 -> 0.43 | Improved |
| Add color statistics to spectrum-only statistical noise (ablation) | AlexNet contrastive learner · ImageNet-100 | top-1 accuracy: 0.31 -> 0.37 | Improved |
| Add WMM to spectrum-and-color statistical noise (ablation) | AlexNet contrastive learner · ImageNet-100 | top-1 accuracy: 0.37 -> 0.40 | Improved |
| Use oriented squares instead of axis-aligned squares in dead-leaves images (ablation) | AlexNet contrastive learner · ImageNet-100 | top-1 accuracy: 0.35 -> 0.37 | Improved |
| Use varied shapes instead of axis-aligned squares in dead-leaves images (ablation) | AlexNet contrastive learner · ImageNet-100 | top-1 accuracy: 0.35 -> 0.39 | Improved |
| Increase perceptual variation between crops beyond the reported sweet spot (ablation) | Contrastive learning on synthetic datasets · ImageNet-100 | top-1 accuracy | Hurt |
Closest papers by meaning
- Procedural Image Programs for Representation Learning (NeurIPS 2022 · similarity 0.87)
- How Learning by Reconstruction Produces Uninformative Features For Perception (ICML 2024 · similarity 0.86)
- Denoising Autoregressive Representation Learning (ICML 2024 · similarity 0.86)
- The Unmet Promise of Synthetic Training Images: Using Retrieved Real Images Performs Better (NeurIPS 2024 · similarity 0.86)
- How Much is a Noisy Image Worth? Data Scaling Laws for Ambient Diffusion. (ICLR 2025 · similarity 0.86)
- Separating Knowledge and Perception with Procedural Data (ICML 2025 · similarity 0.85)
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.
