Research map / Image generation / Autoregressive and latent generation
PUG: Photorealistic and Semantically Controllable Synthetic Data for Representation Learning
Florian Bordes, Shashank Shekhar, Mark Ibrahim, Diane Bouchacourt, Pascal Vincent, Ari S. Morcos
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 |
|---|---|---|---|
| fine-tune CLIP on PUG:AR4T (200K) | CLIP-ViT-B/32 · ARO average | accuracy: 57.32 -> 65.36 | Improved |
| fine-tune CLIP on PUG:AR4T (1M instead of 200K) | CLIP-ViT-B/32 · ARO average | accuracy: 65.36 -> 67.52 | Improved |
| fine-tune CLIP on PUG:AR4T (200K) | CLIP-ViT-B/32 · PUG: SPAR single animal left/right | Precision@1: 49.84 -> 50.78 | No clear effect |
| fine-tune CLIP on PUG:AR4T (200K) | CLIP-ViT-B/32 · PUG: SPAR double animal left/right | Precision@1: 54.42 -> 54.23 | No clear effect |
| fine-tune CLIP on PUG:AR4T (1M instead of 200K) | CLIP-ViT-B/32 · PUG: SPAR | Precision@1: 50.78 -> 50.16 | No clear effect |
| fine-tune CLIP on PUG:AR4T instead of Syn-CLIP | CLIP-ViT-B/32 · ARO average | accuracy: 67.09 -> 65.36 | No clear effect |
| train on 30 backgrounds instead of 50 (ablation) | ResNet50 · PUG: Animals held-out backgrounds | accuracy: 80 -> drops significantly | Hurt |
| show every background for some animals (C>0) (ablation) | ResNet50 · PUG: Animals held-out backgrounds | accuracy | Hurt |
| at least 10 animals seen with every texture (ablation) | ResNet50 · PUG: Animals held-out textures | accuracy | Improved |
| train on medium size only vs small only (ablation) | ResNet50 · PUG: Animals held-out sizes | accuracy | Improved |
| higher ImageNet val accuracy (ViT-B32 21k vs Swin-B) | ViTBasePretrained21k · PUG: ImageNet | top-1 accuracy: 84.3 -> 83.6 | No clear effect |
| hard-negative caption instead of full caption set | VLMs · PUG: SPAR | accuracy | Hurt |
Closest papers by meaning
- Procedural Image Programs for Representation Learning (NeurIPS 2022 · similarity 0.87)
- When Pretty Isn't Useful: Investigating Why Modern Text-to-Image Models Fail as Reliable Training Data Generators (CVPR 2026 · similarity 0.87)
- Free Supervision From Video Games (CVPR 2018 · similarity 0.86)
- Coarse-to-Real: Generative Rendering for Populated Dynamic Scenes (NeurIPS 2026 · similarity 0.86)
- The Unmet Promise of Synthetic Training Images: Using Retrieved Real Images Performs Better (NeurIPS 2024 · similarity 0.85)
- Real-Fake: Effective Training Data Synthesis Through Distribution Matching (ICLR 2024 · 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.
