Research map / Image generation / Text-to-image diffusion
ReGround: Improving Textual and Spatial Grounding at No Cost
Phillip Y. Lee, Minhyuk Sung
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 |
|---|---|---|---|
| rewire gated self-attention and cross-attention from sequential to parallel at inference time | pretrained GLIGEN · MS-COCO-2014 | CLIP score: 30.44 -> 31.29 | Improved |
| rewire to parallel attention (spatial grounding preserved) | pretrained GLIGEN · MS-COCO-2014 | YOLO score: 58.13 -> 56.96 | No clear effect |
| rewire to parallel attention on counting prompts | pretrained GLIGEN · NSR-1K-GPT Counting | CLIP score: 32.46 -> 33.20 | Improved |
| reduce gated self-attention activation to the first 0.1 of steps (scheduled sampling) (ablation) | GLIGEN · MS-COCO-2014 | YOLO score: 58.13 -> 22.75 (CLIP 30.44 -> 31.65) | Hurt |
| reduce gated self-attention activation to the first 0.1 of steps (ablation) | GLIGEN · NSR-1K-GPT Counting | YOLO score: 65.36 -> 26.38 | Hurt |
| remove cross-attention after gated self-attention in GLIGEN (ablation) | GLIGEN · qualitative cat/people examples | spatial grounding | No clear effect |
| rewire to parallel attention when 50% of category boxes are dropped (COCO-Drop) | pretrained GLIGEN · COCO-Drop | CLIP score gap: gap 1.57 times that of original MS-COCO-2017 | Improved |
| rewire GLIGEN to parallel attention | GLIGEN · MS-COCO-2017 | FID | Improved |
| rewire GLIGEN to parallel attention | GLIGEN · user study (92 valid participants) | preference rate (%): 29.95 -> 70.05 | Improved |
| use ReGround rewired GLIGEN as backbone for BoxDiff | BoxDiff · NSR-1K-GPT | CLIP score at same spatial accuracy | Improved |
| apply rewiring to GLIGEN trained on SD v2.1 | GLIGEN (SD v2.1) · MS-COCO | CLIP score vs YOLO score | Improved |
Closest papers by meaning
- GrounDiT: Grounding Diffusion Transformers via Noisy Patch Transplantation (NeurIPS 2024 · similarity 0.88)
- Devil is in the Detail: Towards Injecting Fine Details of Image Prompt in Image Generation via Conflict-free Guidance and Stratified Attention (CVPR 2025 · similarity 0.88)
- R&B: Region and Boundary Aware Zero-shot Grounded Text-to-image Generation (ICLR 2024 · similarity 0.88)
- UGround: Towards Unified Visual Grounding with Unrolled Transformers (ICML 2026 · similarity 0.87)
- Grounded Text-to-Image Synthesis with Attention Refocusing (CVPR 2024 · similarity 0.86)
- Precise Parameter Localization for Textual Generation in Diffusion Models (ICLR 2025 · similarity 0.85)
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.
