Research map / Vision-language and multimodal / Multimodal LLMs
COMPASS: Grounding Composition-Intent Guidance in Unified Multimodal Models
Ziqi Zhou, Weize Quan, Mining Tan, Zhihan Chen, Dandan Zheng, Jingdong Chen, JUN ZHOU, Weiming Dong, Dong-ming Yan
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 composition-specific MoE experts (C-MoE) (ablation) | COMPASS perception · Comp-11 perception test | mAP: 78.2 -> 90.6 | Improved |
| add learnable expert token tau_c with classification supervision (ablation) | COMPASS perception · Comp-11 perception test | mAP: 71.1 -> 90.6 | Improved |
| add cross-attention refinement (ablation) | COMPASS generation · Comp-11 generation | CLIPScore: 0.79 -> 0.83 | Improved |
| add cross-attention refinement (ablation) | COMPASS generation · Comp-11 generation | FID: 8.0 -> 8.2 (slightly worse FID with it); Comp-Cons. 0.64 -> 0.74 | Mixed |
| add customized attention mask (ablation) | COMPASS generation · Comp-11 generation | CLIPScore: 0.72 -> 0.79 | Improved |
| add expert-token conditioning in diffusion timestep embedding (ablation) | COMPASS generation · Comp-11 generation | Comp-Cons.: 0.53 -> 0.59 | Improved |
| replace pixelated grayscale reference with raw reference (ablation) | BAGEL finetune · Comp-11 generation | composition transfer | No clear effect |
| use unified COMPASS vs unified baselines for composition recognition | COMPASS vs BAGEL · Comp-11 perception test | mAP: 63.2 -> 90.6 | Improved |
| compare generation quality FID with Step1X-Edit | COMPASS · Comp-11 generation | FID: 7.5 -> 8.2 | Hurt |
| evaluate zero-shot OOD | COMPASS · UNIAA subset | mAP: 76.2 -> 89.2 vs PhotoFramer | 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.
- PhotoFramer — PhotoFramer: Multi-modal Image Composition Instruction (CVPR 2026) · better on 1 of 1 shared result
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
- OmniGen2: Towards Instruction-Aligned Multimodal Generation (CVPR 2026 · similarity 0.86)
- Think, Plan, Paint: Layout-Aware Reasoning for Controllable Image Generation in Unified Models (NeurIPS 2026 · similarity 0.86)
- Breaking Dual Bottlenecks: Evolving Unified Multimodal Models into Self-Adaptive Interleaved Visual Reasoners (ICML 2026 · similarity 0.85)
- Learning to Compose: Improving Object Centric Learning by Injecting Compositionality (ICLR 2024 · similarity 0.85)
- DecompDreamer: A Composition-Aware Curriculum for Structured 3D Asset Generation (NeurIPS 2026 · similarity 0.85)
- ROVER: Benchmarking Reciprocal Cross-Modal Reasoning for Omnimodal Generation (ICLR 2026 · 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.
