Research map / Image generation / Text-to-image diffusion
CTCal: Rethinking Text-to-Image Diffusion Models via Cross-Timestep Self-Calibration
Xiefan Guo, Xinzhu Ma, Haiyu Zhang, Di Huang
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 |
|---|---|---|---|
| naively align cross-attention maps of all tokens across timesteps (ablation) | SD 2.1 + GORS · T2I-CompBench++ Color (B-VQA) | score: 0.6426 -> 0.6286 | Hurt |
| restrict alignment to noun-token attention maps (ablation) | CTCAL · T2I-CompBench++ Color (B-VQA) | score: 0.6286 -> 0.6897 | Improved |
| add pixel-semantic joint optimization (ablation) | CTCAL · T2I-CompBench++ Color (B-VQA) | score: 0.6897 -> 0.6992 | Improved |
| add subject response alignment regularization (ablation) | CTCAL · T2I-CompBench++ Color (B-VQA) | score: 0.6992 -> 0.7148 | Improved |
| add timestep-aware adaptive weighting (ablation) | CTCAL · T2I-CompBench++ Color (B-VQA) | score: 0.7148 -> 0.7233 | Improved |
| add timestep-aware adaptive weighting (ablation) | CTCAL · T2I-CompBench++ 2D-Spatial (UniDet) | score: 0.2095 -> 0.2142 | Improved |
| fix teacher timestep t_tea=0 rather than random sampling (ablation) | CTCAL on SD 2.1 · T2I-CompBench++ Color (B-VQA) | score: 0.7028 -> 0.7233 | Improved |
| also align adjective tokens (ablation) | CTCAL on SD 2.1 · T2I-CompBench++ Texture (B-VQA) | score: 0.6754 -> 0.6877 | Improved |
| apply CTCAL fine-tuning to SD 3 (2B) | SD 3 (2B) · GenEval Overall | score: 0.62 -> 0.69 | Improved |
| apply CTCAL fine-tuning to SD 3 (2B) | SD 3 (2B) · T2I-CompBench++ Shape (B-VQA) | score: 0.5885 -> 0.5968 | Improved |
| apply CTCAL fine-tuning to SD 2.1 vs GORS | SD 2.1 · T2I-CompBench++ Complex (3-in-1) | score: 0.3386 -> 0.3403 | No clear effect |
| apply CTCAL fine-tuning | SD 2.1 · Diversity (Mean LPIPS, Color) | Mean LPIPS: 0.637 -> 0.634 | No clear effect |
Closest papers by meaning
- Improving Long-Text Alignment for Text-to-Image Diffusion Models (ICLR 2025 · similarity 0.90)
- Diffusion Adaptive Text Embedding for Text-to-Image Diffusion Models (NeurIPS 2025 · similarity 0.90)
- Aligning Text to Image in Diffusion Models is Easier Than You Think (NeurIPS 2025 · similarity 0.90)
- Information Theoretic Text-to-Image Alignment (ICLR 2025 · similarity 0.89)
- CoMat: Aligning Text-to-Image Diffusion Model with Image-to-Text Concept Matching (NeurIPS 2024 · similarity 0.89)
- Noise Diffusion for Enhancing Semantic Faithfulness in Text-to-Image Synthesis (CVPR 2025 · similarity 0.89)
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.
