Research map / Image generation / Text-to-image diffusion
Revisiting Global Text Conditioning in Diffusion Transformers
Nikita Starodubcev, Daniil Pakhomov, Zongze Wu, Ilya Drobyshevskiy, Yuchen Liu, Zhonghao Wang, Yuqian Zhou, Zhe Lin, Dmitry Baranchuk
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 |
|---|---|---|---|
| remove pooled CLIP embedding (set to zero) for long prompts (ablation) | FLUX schnell · MJHQ long prompts | ImageReward: 10.3 -> 10.4 | No clear effect |
| remove pooled CLIP embedding for short prompts (ablation) | FLUX schnell · MJHQ short prompts | ImageReward: 6.2 -> 4.5 | Improved |
| remove pooled CLIP embedding (ablation) | HiDream-Fast · MJHQ short and long prompts | ImageReward: 7.9 -> 8.1 (short), 12.8 -> 13.0 (long) | No clear effect |
| add pooled CLIP embedding to CLIP-free COSMOS without guidance (ablation) | COSMOS · Side-by-side complexity win rate | complexity win rate (%): 43 | Hurt |
| add modulation guidance (aesthetics) | FLUX schnell · COCO 5k | ImageReward: 10.2 -> 11.0 | Improved |
| add modulation guidance | FLUX schnell · GenEval object counting | score: 56 -> 65 | Improved |
| add modulation guidance on CausVid | CausVid 1.3B · VBench | dynamic degree: 75.25 -> 86.59 | Improved |
| add pooled CLIP embedding without guidance (ablation) | CausVid 1.3B · VBench | total score: 62.72 -> 62.82 | No clear effect |
| use dynamic layer-wise guidance instead of constant scale (ablation) | modulation guidance on FLUX schnell · hands correction, human SbS | win rate (%): 48 -> 59 | Improved |
| use dynamic step guidance instead of constant scale (ablation) | modulation guidance on FLUX schnell · aesthetics, human SbS | win rate (%): 62 -> 72 | Improved |
| use modulation guidance instead of Normalized Attention Guidance | FLUX schnell · aesthetics, human SbS | win rate (%): 33 -> 67 | Improved |
| use modulation guidance instead of Concept Sliders | FLUX schnell · hands defects, human SbS | win rate (%): 42 -> 58 | Improved |
Closest papers by meaning
- On Discrete Prompt Optimization for Diffusion Models (ICML 2024 · similarity 0.91)
- Rare Text Semantics Were Always There in Your Diffusion Transformer (NeurIPS 2025 · similarity 0.90)
- Text Diffusion with Reinforced Conditioning (AAAI 2024 · similarity 0.89)
- Exploring the Role of Large Language Models in Prompt Encoding for Diffusion Models (NeurIPS 2024 · similarity 0.89)
- Precise Parameter Localization for Textual Generation in Diffusion Models (ICLR 2025 · similarity 0.89)
- Enhancing Text-to-Image Diffusion Transformer via Split-Text Conditioning (NeurIPS 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.
