Research map / Image generation / Text-to-image diffusion
MC^2: Multi-concept Guidance for Customized Multi-concept Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MC2, an inference-time multi-concept guidance method that composes separately trained heterogeneous single-concept customized models (Textual Inversion, LoRA, DreamBooth) by refining cross-attention maps, and introduces the MC++ benchmark.
Problem
Composing multiple separately trained single-concept customized text-to-image models without access to their training data, joint training, or extra layout conditions, while avoiding attribute leakage between concepts.
Approach
Run parallel diffusion UNets, each with one customized module, on the same noisy latent; at each step, optimize the latent with a loss that encourages cross-attention regions of different concepts to be spatially disjoint and trigger tokens of a concept to overlap (IoU), then merge the denoised outputs with concept-specific weights (a classifier-free-guidance-like composition).
Key results
- On MC++ (customized multi-concept), MC2 achieves overall CLIP-T 0.803 vs. 0.802 for Cones 2 and 0.731 for Mix-of-Show, with the highest CLIP-I (0.658) and DINO (0.336) overall.
- On compositional generation, MC2 reaches Full Prompt 0.356/0.374 and Text-Text 0.833/0.802 (two/three objects), outperforming Attend-and-Excite and Divide-and-Bind.
- In a 28-user study, MC2 was preferred for text alignment in 78.75% (customization) and 82.26% (compositional) of cases.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MC2 instead of Cones 2 | Stable Diffusion v1-5 customized multi-concept generation · MC++ overall | CLIP-T: 0.802 -> 0.803 | Helped |
| use MC2 instead of Mix-of-Show | Stable Diffusion v1-5 customized multi-concept generation · MC++ overall | CLIP-I: 0.647/0.634 -> 0.658/0.638 | Helped |
| use MC2 instead of Mix-of-Show | Stable Diffusion v1-5 customized multi-concept generation · MC++ overall | DINO: 0.334/0.286 -> 0.336/0.287 | Helped |
| use MC2 instead of Attend-and-Excite | Stable Diffusion v1-5 compositional generation · Two objects | Full Prompt Similarity: 0.352 -> 0.356 | Helped |
| use MC2 instead of Attend-and-Excite | Stable Diffusion v1-5 compositional generation · Two objects | Minimum Object Similarity: 0.264 -> 0.266 | Helped |
| use MC2 instead of CONFORM | Stable Diffusion v1-5 compositional generation · Two objects | Text-Text Similarity: 0.824 -> 0.833 | Helped |
| use MC2 instead of Divide-and-Bind | Stable Diffusion v1-5 compositional generation · Three objects | Full Prompt Similarity: 0.372 -> 0.374 | Helped |
| use MC2 instead of Attend-and-Excite | Stable Diffusion v1-5 compositional generation · Three objects | Minimum Object Similarity: 0.233 -> 0.235 | Helped |
| use MC2 instead of CONFORM | Stable Diffusion v1-5 compositional generation · Three objects | Text-Text Similarity: 0.816 -> 0.802 | Hurt |
| add attention grounding training (ablation) | MC2 customized multi-concept generation · MC++ ablation subset | CLIP-T: 0.784 -> 0.789 | Helped |
| add intra-prompt aggregation loss (ablation) | MC2 customized multi-concept generation · MC++ ablation subset | CLIP-T: 0.779 -> 0.789 | Helped |
| add inter-prompt disentanglement loss (ablation) | MC2 customized multi-concept generation · MC++ ablation subset | CLIP-T: 0.440 -> 0.789 | Helped |
Limitations
- Fails to compose two people; attention maps are insufficient to distinguish identities, requiring an extra refinement stage not included in the main method.
- Inference is slower due to parallel UNets and per-step optimization over the latent (not quantified in main text).
- Relies on trigger-token cross-attention maps, which can be messy for some concepts without attention grounding training.
Experimental setup
Datasets: CustomConcept101 Attend-and-Excite compositional prompt benchmark
Benchmarks: MC++ Attend-and-Excite compositional prompt set
Metrics: CLIP-T CLIP-I DINO Full Prompt Similarity Minimum Object Similarity Text-Text Similarity user study preference
Baselines: Custom Diffusion Mix-of-Show Cones 2 Textual Inversion MoLE Attend-and-Excite Divide-and-Bind CONFORM Composable Diffusion
Builds on: Composable Diffusion Attend-and-Excite Custom Diffusion DreamBooth Textual Inversion
Closest papers by meaning
- Concept Weaver: Enabling Multi-Concept Fusion in Text-to-Image Models (CVPR 2024 · similarity 0.91)
- FreeCustom: Tuning-Free Customized Image Generation for Multi-Concept Composition (CVPR 2024 · similarity 0.91)
- Concept Conductor: Orchestrating Multiple Personalized Concepts in Text-to-Image Synthesis (AAAI 2025 · similarity 0.90)
- LaTexBlend: Scaling Multi-concept Customized Generation with Latent Textual Blending (CVPR 2025 · similarity 0.90)
- Attention Calibration for Disentangled Text-to-Image Personalization (CVPR 2024 · similarity 0.89)
- MultiBooth: Towards Generating All Your Concepts in an Image from Text (AAAI 2025 · similarity 0.89)
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.