Research map / Image generation / Text-to-image diffusion
VSC: Visual Search Compositional Text-to-Image Diffusion Model
Do Huu Dat, Nam Hyeon-Woo, Po-Yuan Mao, Tae-Hyun Oh
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 |
|---|---|---|---|
| add VSC visual prototype fusion | SD3.5 · T2I-CompBench | BLIP-VQA harmonic mean: 0.639 -> 0.727 | Improved |
| add VSC visual prototype fusion | SD2.1 · T2I-CompBench | BLIP-VQA harmonic mean: 0.467 -> 0.608 | Improved |
| add VSC visual prototype fusion | SD1.4 · T2I-CompBench | BLIP-VQA harmonic mean: 0.375 -> 0.568 | Improved |
| use VSC instead of SynGen with five binding pairs | SD2.1 · Prompts with 5 binding pairs | BLIP-VQA score: 0.1189 -> 0.1901 | Improved |
| use VSC instead of SynGen with four binding pairs | SD3.5 · Prompts with 4 binding pairs | BLIP-VQA score: 0.5664 -> 0.6183 | Improved |
| use VSC instead of SynGen with three binding pairs | SD3.5 · Prompts with 3 binding pairs | BLIP-VQA score: 0.8081 -> 0.8132 | No clear effect |
| train only on color category instead of all categories (ablation) | VSC SD3.5 · T2I-CompBench color | BLIP-VQA score: 0.85 -> 0.89 | Improved |
| train only on color category instead of all categories (ablation) | VSC SD3.5 · T2I-CompBench texture | BLIP-VQA score: 0.79 -> 0.74 | Hurt |
| train only on color category (transfer to shape) versus untrained baseline (ablation) | VSC SD2.1 · T2I-CompBench shape | BLIP-VQA score: 0.42 -> 0.46 | Improved |
| scale synthetic training set from 5k to 30k images (ablation) | VSC SD2.1 and SD3.5 · T2I-CompBench | BLIP-VQA score: 5,000 to 30,000 images | Improved |
| scale synthetic training set beyond 30k to 90k images (ablation) | VSC SD2.1 and SD3.5 · T2I-CompBench | BLIP-VQA score: 30,000 to 90,000 images | No clear effect |
| train last 2-3 layers of image encoder versus frozen encoder (ablation) | VSC · Localization loss | localization loss: converges to approximately -0.04 | 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.
- InitNO — InitNO: Boosting Text-to-Image Diffusion Models via Initial Noise Optimization (CVPR 2024) · 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
- ORCA: Hunting Compositional Failures in Text-to-Image Diffusion (NeurIPS 2026 · similarity 0.91)
- Magnet: We Never Know How Text-to-Image Diffusion Models Work, Until We Learn How Vision-Language Models Function (NeurIPS 2024 · similarity 0.91)
- Compositional Text-to-Image Synthesis with Attention Map Control of Diffusion Models (AAAI 2024 · similarity 0.90)
- Progressive Compositionality in Text-to-Image Generative Models (ICLR 2025 · similarity 0.90)
- CONFORM: Contrast is All You Need for High-Fidelity Text-to-Image Diffusion Models (CVPR 2024 · similarity 0.90)
- LLM Blueprint: Enabling Text-to-Image Generation with Complex and Detailed Prompts (ICLR 2024 · similarity 0.90)
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.
