Research map / Image generation / Autoregressive and latent generation
Semantic Jitter: Dense Supervision for Visual Comparisons via Synthetic Images
Aron Yu, Kristen Grauman
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 human-verified semantic jitter pairs (DSynth) to real training pairs (ablation) | DeepSTN · PFSmile | accuracy: 81.52 -> 85.78 | Improved |
| add automatically labeled semantic jitter pairs (DSynth-Auto) (ablation) | RankSVM · PFSmile | accuracy: 69.29 -> 75.00 | Improved |
| add automatically labeled semantic jitter pairs (DSynth-Auto) (ablation) | DeepSTN · PFSmile | accuracy: 81.52 -> 84.36 | Improved |
| use semantic jitter instead of low-level jitter (ablation) | DeepSTN · PFSmile | accuracy: 80.09 -> 85.78 | Improved |
| add low-level jitter to real pairs (ablation) | DeepSTN · PFSmile | accuracy: 81.52 -> 80.09 | No clear effect |
| add pseudo-pairs from generator attribute strengths (Real+) (ablation) | DeepSTN · PFSmile | accuracy: 81.52 -> 80.84 | No clear effect |
| add pseudo-pairs from generator attribute strengths (Real+) (ablation) | RankSVM · PFSmile | accuracy: 69.29 -> 68.95 | No clear effect |
| replace human-verified (DSynth) with auto-labeled (DSynth-Auto) synthetic pairs (ablation) | DeepSTN · UT-Zap50K lexicon, Comfort | accuracy: 82.18 -> 87.27 | Improved |
| add DSynth-Auto synthetic pairs (ablation) | DeepSTN · UT-Zap50K lexicon, Comfort | accuracy: 84.95 -> 87.27 | Improved |
| add low-level jitter to real pairs (ablation) | DeepSTN · UT-Zap50K lexicon, Comfort | accuracy: 84.95 -> 81.94 | Hurt |
| add human-verified synthetic pairs (DSynth) (ablation) | DeepSTN · Synthetic test pairs, Sleek | accuracy: 78.76 -> 98.12 | Improved |
| replace DeepSTN baseline with DeepSTN trained with DSynth-Auto | DeepSTN · UT-Zap50K-2 (Open) | accuracy: 70.73 -> 72.18 | Improved |
Closest papers by meaning
- Thinking Outside the Pool: Active Training Image Creation for Relative Attributes (CVPR 2019 · similarity 0.88)
- An Image is Worth More Than a Thousand Words: Towards Disentanglement in The Wild (NeurIPS 2021 · similarity 0.87)
- The Unmet Promise of Synthetic Training Images: Using Retrieved Real Images Performs Better (NeurIPS 2024 · similarity 0.87)
- Efficient Discovery and Effective Evaluation of Visual Perceptual Similarity: A Benchmark and Beyond (ICCV 2023 · similarity 0.87)
- Benchmarking and Analyzing Generative Data for Visual Recognition (TPAMI 2025 · similarity 0.86)
- Scaling Laws of Synthetic Images for Model Training ... for Now (CVPR 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.
