Research map / Image generation / Text-to-image diffusion
StructMAR: Structure-Aware Masked Autoregression for Explicit Layout Alignment in Text-to-Image Generation
Gang Cao, Junying Zhang
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 layout concatenation to plain MAR (ablation) | Plain MAR · COCO-Position | AP: 15.20 -> 22.30 | Improved |
| use layout cross-attention instead of concatenation (ablation) | MAR · COCO-Position | AP: 22.30 -> 27.60 | Improved |
| add soft layout attention bias (ablation) | MAR · COCO-Position | AP: 27.60 -> 45.80 | Improved |
| make the X->L layout bias binary instead of soft (ablation) | MAR · COCO-Position | AP: 45.80 -> 51.00 | Improved |
| add 2D RoPE and gated cross-attention to the binary bias model (ablation) | StructMAR · COCO-Position | AP: 51.00 -> 52.40 | Improved |
| add GRPO fine-tuning (ablation) | StructMAR · COCO-Position | AP: 52.40 -> 57.20 | Improved |
| add GRPO fine-tuning (ablation) | StructMAR · COCO-Position | mIoU: 75.90 -> 79.40 | Improved |
| use GroundingDINO as GRPO reward detector instead of Faster R-CNN (ablation) | StructMAR · COCO-Position | AP: 57.20 -> 57.68 | Improved |
| use YOLOv8 instead of Faster R-CNN as reward detector (ablation) | StructMAR · COCO-Position | AP: 57.20 -> 56.14 | Hurt |
| add GRPO with Faster R-CNN reward (ablation) | StructMAR · COCO-MIG | instance success rate: 58.58 -> 61.70 | Improved |
| replace diffusion MIGC with StructMAR | layout-to-image generator · COCO-MIG (L2) | instance success rate: 67.70 -> 66.15 | Hurt |
| replace diffusion MIGC with StructMAR | layout-to-image generator · COCO-MIG (L6) | instance success rate: 56.88 -> 62.80 | 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.
- MIGC — MIGC: Multi-Instance Generation Controller for Text-to-Image Synthesis (CVPR 2024) · better on 7 of 9 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
- InstanceAssemble: Layout-Aware Image Generation via Instance Assembling Attention (NeurIPS 2025 · similarity 0.88)
- OverLayBench: A Benchmark for Layout-to-Image Generation with Dense Overlaps (NeurIPS 2025 · similarity 0.88)
- Check Locate Rectify: A Training-Free Layout Calibration System for Text-to-Image Generation (CVPR 2024 · similarity 0.87)
- SpotActor: Training-Free Layout-Controlled Consistent Image Generation (AAAI 2025 · similarity 0.87)
- Control and Realism: Best of Both Worlds in Layout-to-Image without Training (ICML 2025 · similarity 0.86)
- Rethinking The Training And Evaluation of Rich-Context Layout-to-Image Generation (NeurIPS 2024 · 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.
