Research map / Image generation / Autoregressive and latent generation
Procedural Image Programs for Representation Learning
Manel Baradad, Chun-Fu Chen, Jonas Wulff, Tongzhou Wang, Rogerio Feris, Antonio Torralba, Phillip Isola
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 |
|---|---|---|---|
| mix images from six different shaders with MixUp (ablation) | MoCo v2 pre-training on Shaders-21k · ImageNet-1k | top-1 accuracy: 30.25 -> 44.83 | Improved |
| mix images from six different shaders with MixUp (ablation) | MoCo v2 pre-training on Shaders-21k · VTAB Structured | top-1 accuracy: 32.85 -> 31.84 | Hurt |
| use MixUp rather than original shader samples (ablation) | MoCo v2 pre-training on StyleGAN Oriented · ImageNet-1k | top-1 accuracy: 38.12 -> 31.73 | Hurt |
| render new shader samples during each training batch (ablation) | S-21k MixUp · ImageNet-1k | top-1 accuracy: 44.83 -> 45.25 | Improved |
| use supervised contrastive learning rather than SimCLR (ablation) | Shaders-21k pre-training · ImageNet-100 | top-1 accuracy: 47.74 -> 48.10 | No clear effect |
| increase the number of shader programs (ablation) | representation learning on Shaders-21k · ImageNet-100 | top-1 accuracy | Improved |
| select curated TwiGL shaders rather than a random same-size subset of Shaders-21k (ablation) | SimCLR pre-training · ImageNet-100 | top-1 accuracy: 42.26 -> 43.26 | Improved |
| increase MixUp from one raw sample to five mixed samples (ablation) | S-21k image generation · ImageNet-100 FID | FID: 41.00 -> 35.03 | Improved |
| generate interpolated images with StyleGAN rather than MixUp (ablation) | S-21k image generation · ImageNet-1k | top-1 accuracy: 44.83 -> 35.19 | Hurt |
| add MixUp to StyleGAN-generated S-21k images (ablation) | S-21k StyleGAN · ImageNet-1k | top-1 accuracy: 35.19 -> 36.46 | Improved |
| use supervised classification instead of unsupervised SimCLR (ablation) | representation learning with shader programs · ImageNet-100 | top-1 accuracy: supervision helps with more than 20 and fewer than 1000 shaders; not outside this regime | Mixed |
Closest papers by meaning
- Separating Knowledge and Perception with Procedural Data (ICML 2025 · similarity 0.88)
- The Unmet Promise of Synthetic Training Images: Using Retrieved Real Images Performs Better (NeurIPS 2024 · similarity 0.86)
- Real-Fake: Effective Training Data Synthesis Through Distribution Matching (ICLR 2024 · similarity 0.86)
- Personalized Representation from Personalized Generation (ICLR 2025 · similarity 0.85)
- Can Generative Models Improve Self-Supervised Representation Learning? (AAAI 2025 · similarity 0.85)
- CO-SPY: Combining Semantic and Pixel Features to Detect Synthetic Images by AI (CVPR 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.
