Research map / Image generation / Text-to-image diffusion
Event-Customized Image 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.
Introduces event-customized image generation from a single reference image and proposes FreeEvent, a training-free method combining cross-attention entity switching with spatial/self-attention injection for event transfer on Stable Diffusion.
Problem
Existing customization methods only handle basic single-entity actions or two-entity interactions and need multiple near-identical reference images; complex multi-entity events from a single reference image are not addressed.
Approach
A training-free diffusion framework with two extra paths: an entity switching path that uses mask-guided cross-attention guidance and regulation to place target entities at reference locations, and an event transferring path that injects reference spatial features and self-attention maps into the denoising process.
Key results
- On SWiG-Event, FreeEvent reaches 41.12 R@1 image retrieval versus 10.90 for the best baseline (MIGC), and FID 29.05 versus 49.81.
- FreeEvent gets 34.10 T-1 verb detection accuracy versus 10.66 for ControlNet on SWiG-Event.
- In a 10-expert user study on Real-Event, FreeEvent received 50 votes versus 25 for the next best (ControlNet) across 50 trials.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FreeEvent instead of MIGC | event-customized generation · SWiG-Event | R@1: 10.90 -> 41.12 | Helped |
| Use FreeEvent instead of ControlNet | event-customized generation · SWiG-Event | T-1: 10.66 -> 34.10 | Helped |
| Use FreeEvent instead of MIGC | event-customized generation · SWiG-Event | DreamSim: 0.3772 -> 0.6282 | Helped |
| Use FreeEvent instead of MIGC | event-customized generation · SWiG-Event | FID: 49.81 -> 29.05 | Helped |
| Use FreeEvent instead of ControlNet | event customization · Real-Event | HJ: 25 -> 50 | Helped |
| add cross-attention guidance (ablation) | FreeEvent · Real-Event | CLIP-I: 0.5140 -> 0.5771 | Helped |
| add cross-attention regulation (ablation) | FreeEvent · Real-Event | CLIP-I: 0.5459 -> 0.5771 | Helped |
| add event transferring injection (ablation) | FreeEvent · Real-Event | CLIP-I: 0.4945 -> 0.5771 | Helped |
| add cross-attention guidance (ablation) | FreeEvent · Real-Event | DINO: 0.1058 -> 0.2865 | Helped |
| add cross-attention regulation (ablation) | FreeEvent · Real-Event | DINO: 0.1297 -> 0.2865 | Helped |
| add event transferring injection (ablation) | FreeEvent · Real-Event | DINO: 0.0825 -> 0.2865 | Helped |
| add cross-attention guidance (ablation) | FreeEvent · Real-Event | DreamSim: 0.2079 -> 0.3877 | Helped |
Surprising findings
- Training-based customization methods (DreamBooth, ReVersion) fail with only one reference image, showing that single-image event customization needs a training-free approach.
Limitations
- Performance degrades when there are many entities, especially small ones.
- Depends on Stable Diffusion's generative capability, struggling with faces and hands.
- Large shape discrepancy between target and reference entities can leak reference layout into target appearance.
Experimental setup
Datasets: SWiG HICO-DET
Benchmarks: SWiG-Event Real-Event
Metrics: R@1 R@5 R@10 verb detection top-k accuracy CLIP-I DINO DreamSim CLIP-T FID human judgment
Baselines: ControlNet MIGC BoxDiff PnP MAG-Edit DreamBooth ReVersion
Builds on: Stable Diffusion Plug-and-Play BoxDiff MAG-Edit
Closest papers by meaning
- Diffusion-based Event Generation for High-Quality Image Deblurring (CVPR 2025 · similarity 0.84)
- Multi-party Collaborative Attention Control for Image Customization (CVPR 2025 · similarity 0.84)
- EventGPT: Event Stream Understanding with Multimodal Large Language Models (CVPR 2025 · similarity 0.83)
- CustomCrafter: Customized Video Generation with Preserving Motion and Concept Composition Abilities (AAAI 2025 · similarity 0.83)
- VideoMage: Multi-Subject and Motion Customization of Text-to-Video Diffusion Models (CVPR 2025 · similarity 0.83)
- EF-3DGS: Event-Aided Free-Trajectory 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.83)
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.