Research map / Image generation / Text-to-image diffusion
From Structural Feedback to Prompt Policies: Learning Faithful Text-to-Image Prompt Editors
Wentao Ye, Yali Ye, Zhiqing Xiao, Ru Peng, Xinpeng Ti, Liyao Li, Zhanming Shen, Peng Lu, Jie Zhang, Haobo Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes learning faithful text-to-image prompt editors that edit user prompts guided by structural feedback rather than open-ended rewriting that may drift from user intent.
Problem
Text-to-image prompt optimization by open-ended rewriting can silently weaken the user's intent, especially for prompts specifying fine-grained objects, attributes, and relations.
Approach
Reframes prompt optimization as faithful prompt editing that learns a prompt policy from structural feedback on whether the user's specified objects, attributes, and relations are preserved in generated images.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add DSE scene-graph calibration (ablation) | StruPO · Lexica | SGR (%): 73.69 -> 80.74 | Improved |
| add graph encoder (MHE) (ablation) | StruPO · Lexica | SGR (%): 75.33 -> 80.74 | Improved |
| use edit actions (delete/reorder/insert) instead of free-form generation (ablation) | StruPO · Lexica | SGR (%): 73.68 -> 80.74 | Improved |
| add image-grounded RL after SFT (ablation) | StruPO · Lexica | SGR (%): 74.04 -> 80.74 | Improved |
| add structured consistency reward r_con (ablation) | StruPO · Lexica | SGR (%): 75.14 -> 80.74 | Improved |
| add structured consistency reward r_con (ablation) | StruPO · Lexica | AES_delta: .68 -> .63 | Hurt |
| add RL after SFT (ablation) | StruPO · Lexica | AES_delta: .57 -> .63 | Improved |
| add DSE calibration (ablation) | StruPO · Lexica | AES_delta: .64 -> .63 | No clear effect |
| StruPO vs best baseline on relation recall | SD v1.5 · Lexica | RR (%): 7.96 points over best listed baseline | Improved |
| apply StruPO editor trained on SD v1.5 to unseen generators without retraining | StruPO · Lexica (SD v1.4, SD v2.1, FLUX.1-dev) | RR / SGR: FLUX.1-dev RR 62.43 vs Promptist 53.41 | Improved |
| StruPO versus VisualPrompter on out-of-domain prompts | StruPO · COCO | SGR (%): 78.94 vs 83.26 | Hurt |
| StruPO versus VisualPrompter on out-of-domain prompts | StruPO · COCO | AES_delta: 1.02 vs -0.36 | Improved |
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.
- VisualPrompter — VisualPrompter: Semantic-Aware Prompt Optimization with Visual Feedback for Text-to-Image Synthesis (ICLR 2026) · better on 2 of 3 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
- Textualize Visual Prompt for Image Editing via Diffusion Bridge (AAAI 2025 · similarity 0.89)
- PromptCoT: Align Prompt Distribution via Adapted Chain-of-Thought (CVPR 2024 · similarity 0.89)
- A User-Friendly Framework for Generating Model-Preferred Prompts in Text-to-Image Synthesis (AAAI 2024 · similarity 0.89)
- Prompting Hard or Hardly Prompting: Prompt Inversion for Text-to-Image Diffusion Models (CVPR 2024 · similarity 0.88)
- Multimodal Large Language Models Make Text-to-Image Generative Models Align Better (NeurIPS 2024 · similarity 0.88)
- Learning to Sample Effective and Diverse Prompts for Text-to-Image Generation (CVPR 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.
