Research map / Image generation / Text-to-image diffusion
When Pretty Isn't Useful: Investigating Why Modern Text-to-Image Models Fail as Reliable Training Data Generators
Krzysztof Adamkiewicz, Brian B. Moser, Stanislav Frolov, Tobias Christian Nauen, Federico Raue, Andreas Dengel
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 |
|---|---|---|---|
| use newer T2I model as synthetic data generator (class-name prompts) | ResNet-50 trained on synthetic ImageNet subset · ImageNet-1k val (200-class subset) | Synth-to-Real accuracy | Hurt |
| use higher GenEval / CLIPScore T2I model (class-name prompts) | synthetic-data classifiers (ResNet-50, ViT-Ti, ConvNeXt-Ti, Swin-Ti) · ImageNet-1k val (200-class subset) | Synth-to-Real accuracy | Hurt |
| use detailed captions instead of class names as prompts (ablation) | T2I synthetic data generation · ImageNet-1k val (200-class subset) | Synth-to-Real accuracy, coverage | Improved |
| use detailed captions for texture-space classifier (ablation) | BagNet texture classifier on synthetic data · ImageNet-1k val (200-class subset) | accuracy | No clear effect |
| remove texture by training in depth (structure) space (ablation) | ResNet-50 on synthetic data · ImageNet-1k val (200-class subset) | Synth-to-Real gap | Improved |
| low-pass filter instead of high-pass filter images (ablation) | ResNet-50 on synthetic data · ImageNet-1k val (200-class subset) | Synth-to-Real gap | Improved |
| detailed captions for high-frequency filtered data (ablation) | ResNet-50 on synthetic data · ImageNet-1k val (200-class subset) | accuracy | No clear effect |
| detailed captions instead of class names (ablation) | Synth-to-Real vs Real-to-Synth asymmetry · ImageNet-1k val (200-class subset) | transfer asymmetry | Improved |
Closest papers by meaning
- The Unmet Promise of Synthetic Training Images: Using Retrieved Real Images Performs Better (NeurIPS 2024 · similarity 0.91)
- Scaling Laws of Synthetic Images for Model Training ... for Now (CVPR 2024 · similarity 0.90)
- Real-Fake: Effective Training Data Synthesis Through Distribution Matching (ICLR 2024 · similarity 0.89)
- Learning Visual Generative Priors without Text (CVPR 2025 · similarity 0.88)
- FakeInversion: Learning to Detect Images from Unseen Text-to-Image Models by Inverting Stable Diffusion (CVPR 2024 · similarity 0.88)
- CommonCanvas: Open Diffusion Models Trained on Creative-Commons Images (CVPR 2024 · similarity 0.87)
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.
