Research map / Detection and segmentation / Open-set and few-shot detection
CLiC: Concept Learning in Context
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.
Learns a local visual concept token from one image in the context of its parent object (soft-masked diffusion loss) and transfers it to target objects or generates new objects with it.
Problem
Learning a local visual pattern from a single image and placing it at a specific location on objects in other images (with different orientations and shapes) or generating new objects with it, which naive personalization methods fail to do because they learn the concept in isolation.
Approach
Learn the concept in-context by computing diffusion losses under a soft mask that includes both the in-mask concept and surrounding object area, combined with cross-attention loss on the token, an RoI-only concept loss, and cross-attention guidance plus blended diffusion for placing the concept in target images; also uses diffusion-based automatic RoI matching.
Key results
- In a user study with 42 participants over 30 source-target pairs, CLiC received an average ranking of 3.43 versus 2.33 for RealFill, 2.27 for Break-A-Scene, and 1.96 for Custom Diffusion.
- Concept learning takes about 3 minutes (500 steps) on a single NVIDIA RTX 3090 GPU.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CLiC instead of RealFill | Concept transfer · User study | Average Ranking: 2.33 -> 3.43 | Helped |
| Add the RoI loss (ablation) | CLiC · Concept transfer | concept geometric and structural detail | Helped |
| Add the attention loss (ablation) | CLiC · Concept transfer | edit localization | Helped |
| Add the context loss (ablation) | CLiC · Concept transfer | geometric detail and edit localization | Helped |
| Increase cross-attention guidance step size η (ablation) | CLiC · Concept transfer | concept presence in the target image | Helped |
Limitations
- Performance degrades when source and target domains differ substantially.
- Optimization is time-consuming and not suitable for real-time applications.
- Evaluation is mainly qualitative plus a user study with a limited set of 30 pairs.
Experimental setup
Metrics: user study ranking (average rank, 1-4 scale)
Baselines: Custom Diffusion Break-A-Scene RealFill
Builds on: Custom Diffusion Textual Inversion Blended Diffusion Attend-and-Excite
Closest papers by meaning
- ConceptBed: Evaluating Concept Learning Abilities of Text-to-Image Diffusion Models (AAAI 2024 · similarity 0.86)
- ICE: Intrinsic Concept Extraction from a Single Image via Diffusion Models (CVPR 2025 · similarity 0.85)
- Learning Vision and Language Concepts for Controllable Image Generation (ICML 2025 · similarity 0.85)
- Language-Informed Visual Concept Learning (ICLR 2024 · similarity 0.85)
- An Image is Worth Multiple Words: Discovering Object Level Concepts using Multi-Concept Prompt Learning (ICML 2024 · similarity 0.85)
- Bridging the gap to real-world language-grounded visual concept learning (NeurIPS 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.