Research map / Image generation / Text-to-image diffusion
Extending One-Step Image Generation from Class Labels to Text via Discriminative Text Representation
Chenxi Zhao, Chen Zhu, Xiaokun Feng, Aiming Hao, Jiashu Zhu, Jiachen Lei, Jiahong Wu, Xiangxiang Chu, Jufeng Yang
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 |
|---|---|---|---|
| train with MeanFlow objective (EMF) instead of standard flow matching, 1-step | BLIP3o-NEXT · GenEval overall | GenEval score: 0.46 -> 0.74 | Improved |
| train with MeanFlow (EMF), 4-step | BLIP3o-NEXT · GenEval overall | GenEval score: 0.86 -> 0.90 | Improved |
| train with MeanFlow (EMF), 1-step | BLIP3o-NEXT · DPG-Bench overall | DPG-Bench score: 57.05 -> 77.36 | Improved |
| train with MeanFlow (EMF), 1-step | BLIP3o-NEXT · HPS-v2.1 average | HPS-v2.1 score: 18.54 -> 25.77 | Improved |
| use 4 EMF steps instead of 30 BLIP3o-NEXT steps | EMF · GenEval overall | GenEval score: 0.91 -> 0.90 | No clear effect |
| increase EMF sampling steps from 1 to 4 (ablation) | EMF · GenEval overall | GenEval score: 0.74 -> 0.90 | Improved |
| increase EMF sampling steps from 4 to 8 (ablation) | EMF · DPG-Bench overall | DPG-Bench score: 81.20 -> 81.94 | Improved |
| MeanFlow training with fewer sampling steps (20 -> 4) (ablation) | SANA-1.5 · GenEval overall | GenEval score: 0.83 -> 0.50 | Hurt |
| fine-tune the text encoder on SFT data (ablation) | SANA-1.5 flow matching · GenEval overall | GenEval score: 0.81 -> 0.85 | Improved |
| fine-tune the text encoder on SFT data before MeanFlow, 4 steps (ablation) | SANA-1.5 MeanFlow · GenEval overall | GenEval score: 0.50 -> 0.47 | No clear effect |
| use BLIP3o-NEXT text encoder instead of T5 (ablation) | text-retrieval analysis · COCO2017 retrieval | DINO similarity: 0.634 -> 0.734 | Improved |
| use BLIP3o-NEXT text encoder instead of T5 (ablation) | sub-sequence analysis · DPG-Bench prompts | cosine similarity: 0.893 -> 0.999 | Improved |
Closest papers by meaning
- A Comprehensive Study of Decoder-Only LLMs for Text-to-Image Generation (CVPR 2025 · similarity 0.88)
- LLM4GEN: Leveraging Semantic Representation of LLMs for Text-to-Image Generation (AAAI 2025 · similarity 0.86)
- Elucidating the design space of language models for image generation (ICML 2025 · similarity 0.86)
- LLM Blueprint: Enabling Text-to-Image Generation with Complex and Detailed Prompts (ICLR 2024 · similarity 0.86)
- Discriminative Probing and Tuning for Text-to-Image Generation (CVPR 2024 · similarity 0.86)
- Iterative Self-Training with Class-Aware Text-to-Image Synthesis for Visual Task Learning (AAAI 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.
