Research map / Image generation / Text-to-image diffusion
Evaluating Data Attribution for Text-to-Image Models
Sheng-Yu Wang, Alexei A. Efros, Jun-Yan Zhu, Richard Zhang
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 |
|---|---|---|---|
| tune a linear mapping with contrastive loss on the attribution dataset | DINO, CLIP, ViT, MoCo features · Object-centric and style-centric test sets | Recall@10 | Improved |
| train the mapper on object and style data together rather than one (ablation) | Attribution feature mapper · Object and style axes | Recall@10 | Improved |
| add regularization on the affine mapping (ablation) | Attribution feature mapper · All test cases averaged | Recall@10 | Improved |
| use linear mapping instead of other mapping functions (ablation) | Attribution feature mapper · All test cases averaged | Recall@10 | Improved |
| change augmentation strength (mild vs strong) (ablation) | Attribution feature mapper training · All test cases averaged | Recall@10 | No clear effect |
| use copy-detection (SSCD) or style descriptor (ALADIN) features instead of general features | Pretrained feature spaces · Attribution benchmark | Recall@10 | Hurt |
| train on one distribution and test on unseen classes or source | Tuned features · ImageNet-Unseen and Artchive | Recall@10 | Improved |
| calibrate softmax temperature and threshold for soft influence scores (ablation) | Soft attribution scores · Soft influence scores | score variance: variance < 0.0003 % at tau = 1 | Improved |
| increase fine-tuning set size to 1000 images (ablation) | Exemplar-based attribution transfer to MSCOCO · MSCOCO fine-tuned models | correlation with benchmark | Hurt |
| attribute media-prompted images instead of GPT-prompted images | Object-centric attribution · Prompt types | Recall@10 | Hurt |
Closest papers by meaning
- Data Attribution for Text-to-Image Models by Unlearning Synthesized Images (NeurIPS 2024 · similarity 0.89)
- Fast Data Attribution for Text-to-Image Models (NeurIPS 2025 · similarity 0.89)
- Benchmarking the Attribution Quality of Vision Models (NeurIPS 2024 · similarity 0.86)
- Less is More: Fewer Interpretable Region via Submodular Subset Selection (ICLR 2024 · similarity 0.85)
- quanda: An Interpretability Toolkit for Training Data Attribution Evaluation (NeurIPS 2026 · similarity 0.85)
- Final-Model-Only Data Attribution with a Unifying View of Gradient-Based Methods (NeurIPS 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.
