Research map / Image generation / Text-to-image diffusion
LLM Thought Divergence and Convergence for Dialogue-Based Image Generation Control
Hui 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 |
|---|---|---|---|
| use LTDC instead of GILL | GILL · VisDial val | CLIP similarity: 64.50 -> 69.37 | Improved |
| add LTDC divergence and convergence to the vanilla ChatGPT-4o-mini pipeline | Vanilla(SC) · VisDial val | CLIP similarity: 67.41 -> 69.37 | Improved |
| add LTDC to the vanilla Janus-Pro + Deepseek-R1 pipeline | Vanilla(JD) · VisDial val | CLIP similarity: 68.61 -> 70.03 | Improved |
| add LTDC to vanilla Stable Diffusion + ChatGPT-4o-mini | Vanilla(SC) · VisDial val | LPIPS: 71.38 -> 70.71 | No clear effect |
| replace Stable Diffusion by Janus-Pro with Deepseek-R1 (vanilla) | Vanilla(SD) · VisDial val | CLIP similarity: 67.37 -> 68.61 | Improved |
| replace Stable Diffusion by Janus-Pro with ChatGPT-4o-mini (vanilla) (ablation) | Vanilla(SC) · VisDial val | CLIP similarity: 67.41 -> 67.09 | No clear effect |
| use Deepseek-R1 instead of Llama3.3 for task breakdown (Janus-Pro) (ablation) | Vanilla(JL) · VisDial val | CLIP similarity: 68.30 -> 68.61 | Improved |
| use ChatGPT-4o-latest instead of ChatGPT-4o-mini as evaluator (ablation) | LTDC evaluator · Multi-modal evaluator accuracy (100 VisDial examples) | accuracy (%): 73.82 -> 81.19 | Improved |
| use Llama3.2 (90b) instead of Qwen-vl-plus (7b) as evaluator (ablation) | LTDC evaluator · Multi-modal evaluator accuracy (100 VisDial examples) | accuracy (%): 66.13 -> 77.56 | Improved |
| use ChatGPT-4o as evaluator under the same prompt (ablation) | LTDC evaluator · Multi-modal evaluator accuracy (100 VisDial examples) | accuracy (%): Fail | Hurt |
| use LTDC instead of GILL in human evaluation | GILL · VisDial val (100 samples, 3 human evaluators) | correct ratio (%): 58.3 -> 72.9 | 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.
- GLIDE — GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models (ICML 2022) · better on 2 of 2 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
- ThinkGen: Generalized Thinking for Visual Generation (CVPR 2026 · similarity 0.88)
- Janus-Pro-R1: Advancing Collaborative Visual Comprehension and Generation via Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
- Elucidating the design space of language models for image generation (ICML 2025 · similarity 0.87)
- MindOmni: Unleashing Reasoning Generation in Vision Language Models with RGPO (NeurIPS 2025 · similarity 0.87)
- Thinking-while-Generating: Interleaving Textual Reasoning throughout Visual Generation (CVPR 2026 · similarity 0.87)
- IDEA-Bench: How Far are Generative Models from Professional Designing? (CVPR 2025 · similarity 0.86)
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.
