Research map / Image generation / Text-to-image diffusion
ISAC: Training-Free Instance-to-Semantic Attention Control for Multi-Instance Generation
Sanghyun Jo, Wooyeol Lee, Ziseok Lee, Jonghyun Choi, Jaesik Park, Kyungsu Kim
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 instance-to-semantic schedule instead of instance-only loss (ablation) | ISAC_LO (SD1.5) · IntraCompBench multi-class | accuracy: 10% -> 36% | Improved |
| add instance-to-semantic schedule instead of semantic-only loss (ablation) | ISAC_LO (SD1.5) · IntraCompBench multi-class | accuracy: 28% -> 36% | Improved |
| use instance-to-semantic schedule instead of fixed 0.5/0.5 balance (ablation) | ISAC_LO (SD1.5) · IntraCompBench multi-class | accuracy: 25% -> 36% | Improved |
| use instance-to-semantic instead of reversed semantic-to-instance schedule (ablation) | ISAC_LO (SD1.5) · IntraCompBench multi-instance | accuracy: 55% -> 69% | Improved |
| add normalized coordinates to SA clustering features (ablation) | ISAC_LO (SD1.5) · IntraCompBench multi-class | accuracy: 21% -> 36% | Improved |
| replace K-means with spectral clustering (ablation) | ISAC_LO · IntraCompBench multi-class | accuracy: 36% -> 35% | No clear effect |
| replace K-means with GMM (latency 505 ms -> 8,313 ms) (ablation) | ISAC_LO · IntraCompBench multi-class | accuracy: 36% -> 36% | No clear effect |
| add ISAC_LS latent selection | Flux.1-dev · IntraCompBench multi-class (avg) | accuracy: 31% -> 51% | Improved |
| add ISAC_LS latent selection | Qwen-Image · IntraCompBench multi-class (avg) | accuracy: 48% -> 56% | Improved |
| add ISAC_LO to few-step model | Z-Image-Turbo · IntraCompBench multi-class (avg) | accuracy: 48% -> 68% | Improved |
| add ISAC_LO to few-step model | Flux.2-klein-4B · IntraCompBench multi-class (avg) | accuracy: 34% -> 54% | Improved |
| add ISAC_LO to GLIGEN layout controller | GLIGEN · HRS-Bench Color | accuracy: 0.307 -> 0.452 | Improved |
Closest papers by meaning
- InstanceDiffusion: Instance-level Control for Image Generation (CVPR 2024 · similarity 0.89)
- ROICtrl: Boosting Instance Control for Visual Generation (CVPR 2025 · similarity 0.87)
- Noise Diffusion for Enhancing Semantic Faithfulness in Text-to-Image Synthesis (CVPR 2025 · similarity 0.87)
- Towards Understanding Cross and Self-Attention in Stable Diffusion for Text-Guided Image Editing (CVPR 2024 · similarity 0.87)
- Compositional Text-to-Image Synthesis with Attention Map Control of Diffusion Models (AAAI 2024 · similarity 0.87)
- Grounded Text-to-Image Synthesis with Attention Refocusing (CVPR 2024 · similarity 0.87)
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.
