Research map / Image generation / Text-to-image diffusion
CO3: CONTRASTING CONCEPTS COMPOSE BETTER
Debottam Dutta, Jianchong Chen, RAJALAXMI RAJAGOPALAN, Yu-Lin Wei, Romit Roy Choudhury
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 CO3 resampling in first three DDIM steps (ablation) | SDXL · Attend-Excite | average score: 0.8435 -> 0.9437 | Improved |
| add CO3 corrector after resampling (ablation) | SDXL + Resampling · Attend-Excite | average score: 0.9437 -> 0.9463 | Improved |
| add closeness-aware weight modulation (ablation) | SDXL + Resampling + Corrector · Attend-Excite | average score: 0.9463 -> 1.0118 | Improved |
| add corrector alone to SDXL (ablation) | SDXL · Attend-Excite | average score: 0.8435 -> 0.9086 | Improved |
| add weight modulation to corrector-only variant (ablation) | SDXL + Corrector · Attend-Excite | average score: 0.9086 -> 0.9797 | Improved |
| apply CO3 instead of composable diffusion on PixArt-Sigma | PixArt + cDiff · Attend-Excite Animal-Object | ImageReward: 0.047 -> 1.770 | Improved |
| add CO3 to PixArt-Sigma | PixArt-Sigma · Attend-Excite Animal-Animal | ImageReward: 1.066 -> 1.380 | Improved |
| increase resampling steps from 3 to 4 (ablation) | CO3 · Attend-Excite Animal-Animal (beta 0.8) | ImageReward: 1.2218 -> 1.1705 | Hurt |
| increase corrector timesteps from 4 to 10 (ablation) | CO3 · Attend-Excite Animal-Animal | ImageReward: 1.1862 -> 1.3149 | Improved |
| increase corrector iterations from 7 to 15 (ablation) | CO3 · Attend-Excite Animal-Animal (lambda 0.9) | ImageReward: 1.2622 -> 1.2151 | Hurt |
| raise affinity decay beta from 0.3 to 1.1 (ablation) | CO3 · Attend-Excite Animal-Animal (num ts 7) | ImageReward: 1.1484 -> 1.2566 | Improved |
| add CO3 at 10 DDIM steps | SDXL · Attend-Excite, 10 steps | average score: 0.6183 -> 0.8037 | Improved |
Closest papers by meaning
- Concept Reachability in Diffusion Models: Beyond Dataset Constraints (ICML 2025 · similarity 0.90)
- CoMat: Aligning Text-to-Image Diffusion Model with Image-to-Text Concept Matching (NeurIPS 2024 · similarity 0.90)
- CONFORM: Contrast is All You Need for High-Fidelity Text-to-Image Diffusion Models (CVPR 2024 · similarity 0.90)
- The Hidden Language of Diffusion Models (ICLR 2024 · similarity 0.89)
- You Can’t Have It Both Ways: Concept Entanglement Limits Diffusion Model Unlearning (NeurIPS 2026 · similarity 0.89)
- Efficient bias mitigation in T2I diffusion models using Concept Graphs (NeurIPS 2026 · 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.
