Research map / Image generation / Text-to-image diffusion
DetailMaster: Can Your Text-to-Image Model Handle Long Prompts?
Qirui Jiao, Daoyuan Chen, Yilun Huang, Xika Lin, Ying Shen, Yaliang Li
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 |
|---|---|---|---|
| raise text-encoder token limit from 77 to 512 without long-prompt training (ablation) | SD-XL · DetailMaster Character Locations | accuracy (%): 9.39 -> 10.45 | Improved |
| add long-prompt training at a 77-token limit (ablation) | SD-XL · DetailMaster Character Locations | accuracy (%): 9.39 -> 10.93 | Improved |
| combine 512-token limit with long-prompt training (ablation) | SD-XL · DetailMaster Character Locations | accuracy (%): 9.39 -> 12.52 | Improved |
| combine 512-token limit with long-prompt training (ablation) | SD-XL · DetailMaster Entity Relationships | accuracy (%): 8.43 -> 13.07 | Improved |
| long-prompt training at 77 limit on style attribute versus 512 limit short training (ablation) | SD-XL · DetailMaster Scene Attributes (Style) | accuracy (%): 66.22 -> 53.72 | Hurt |
| use T5 encoder (longer input) instead of CLIP | DeepFloyd IF vs SD1.5/SD-XL · DetailMaster Character Locations | accuracy (%): 10.95 -> 14.28 | Improved |
| add long-prompt training and 512 token limit (ParaDiffusion) versus conventional training at same capacity | DeepFloyd IF · DetailMaster Character Locations | accuracy (%): 14.28 -> 20.32 | Improved |
| add T5 adapter with only 128-token training limit (LLM4GEN) | SD1.5 · DetailMaster Character Locations | accuracy (%): 8.68 -> 9.22 | No clear effect |
| add iterative layout refinement (LLM Blueprint) | SD1.5 · DetailMaster Character Locations | accuracy (%): 8.68 -> 18.44 | Improved |
| optimise SD1.5 for long prompts | SD1.5 · DetailMaster Scene Attributes (Style) | accuracy (%): 80.90 -> 50.11 | Hurt |
| increase prompt length | 10 evaluated T2I models · DetailMaster | accuracy | Hurt |
Closest papers by meaning
- LLM Blueprint: Enabling Text-to-Image Generation with Complex and Detailed Prompts (ICLR 2024 · similarity 0.90)
- Prompt-Free Diffusion: Taking "Text" out of Text-to-Image Diffusion Models (CVPR 2024 · similarity 0.89)
- DynT2I-Eval: A Dynamic Evaluation Framework for Text-to-Image Models (NeurIPS 2026 · similarity 0.89)
- FineGRAIN: Evaluating Failure Modes of Text-to-Image Models with Vision Language Model Judges (NeurIPS 2025 · similarity 0.88)
- ConceptMix: A Compositional Image Generation Benchmark with Controllable Difficulty (NeurIPS 2024 · similarity 0.88)
- OneIG-Bench: Omni-dimensional Nuanced Evaluation for Image Generation (NeurIPS 2025 · similarity 0.88)
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.
