Research map / Image generation / Text-to-image diffusion
Attribute-Centric Compositional Text-to-Image Generation
Yuren Cong, Martin Renqiang Min, Li Erran Li, Bodo Rosenhahn, Michael Ying Yang
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 |
|---|---|---|---|
| add independent CLIP-based matching discriminator D_m (ablation) | ACTIG · CUB | FID: 18.3 -> 13.9 | Improved |
| add independent CLIP-based matching discriminator D_m (ablation) | ACTIG · CUB | R-Precision: 15.2 -> 21.8 | Improved |
| add attribute-centric contrastive loss (ablation) | ACTIG · CUB | FID: 13.9 -> 12.8 | Improved |
| add attribute-centric contrastive loss (ablation) | ACTIG · CUB | R-Precision: 21.8 -> 25.6 | Improved |
| add image-free training with attribute-centric feature augmentation (ablation) | ACTIG · CUB | R-Precision: 25.6 -> 28.7 | Improved |
| add image-free training (ablation) | ACTIG · CUB | FID: 12.8 -> 12.5 | No clear effect |
| finetune CLIP with norm penalty instead of pre-trained CLIP (ablation) | ACTIG · CelebA-HQ | R-Precision: 57.1 -> 65.0 | Improved |
| finetune CLIP with norm penalty instead of pre-trained CLIP (ablation) | ACTIG · CUB | R-Precision: 19.5 -> 28.7 | Improved |
| add norm penalty to CLIP finetuning (ablation) | ACTIG · CelebA-HQ | R-Precision: 61.2 -> 65.0 | Improved |
| add norm penalty to CLIP finetuning (ablation) | ACTIG · CUB | FID: 12.3 -> 12.5 | No clear effect |
| increase attribute count in input text from <=2 to 3 | ACTIG · CelebA-HQ (K=100) | R-Precision: 41.3 -> 67.0 | Improved |
| increase attribute count from 5 to >=6 | ACTIG · CelebA-HQ (K=100) | R-Precision: 67.5 -> 63.8 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- StyleT2I — StyleT2I: Toward Compositional and High-Fidelity Text-to-Image Synthesis (CVPR 2022) · better on 6 of 6 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- DomainGallery: Few-shot Domain-driven Image Generation by Attribute-centric Finetuning (NeurIPS 2024 · similarity 0.85)
- Attribute Based Interpretable Evaluation Metrics for Generative Models (ICML 2024 · similarity 0.85)
- Progressive Compositionality in Text-to-Image Generative Models (ICLR 2025 · similarity 0.85)
- ConceptMix: A Compositional Image Generation Benchmark with Controllable Difficulty (NeurIPS 2024 · similarity 0.84)
- Composing Parts for Expressive Object Generation (CVPR 2025 · similarity 0.84)
- Continuous, Subject-Specific Attribute Control in T2I Models by Identifying Semantic Directions (CVPR 2025 · similarity 0.84)
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.
