Research map / Image generation / Editing and style transfer
Does FLUX Already Know How to Perform Physically Plausible Image Composition?
Shilin Lu, Zhuming Lian, Zihan Zhou, Shaocong Zhang, Chen Zhao, Adams Wai-Kin Kong
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 MSA loss (adapter-guided latent optimization) (ablation) | SHINE · DreamEditBench | DINOv2 / DreamSim: DINOv2 0.6745 -> 0.7204; DreamSim 0.5233 -> 0.3951 | Improved |
| add DSG (blur Q_img negative guidance) (ablation) | SHINE · DreamEditBench | ImageReward / VisionReward: IR 0.5577 -> 0.5633; VR 3.5997 -> 3.6130 | Improved |
| add DSG on top of MSA (ablation) | SHINE · DreamEditBench | ImageReward: 0.5455 -> 0.5745 | Improved |
| add ABB (adaptive background blending) (ablation) | SHINE · DreamEditBench | LPIPS / SSIM: LPIPS 0.0166 -> 0.0210; SSIM 0.9076 -> 0.9010 | No clear effect |
| add ABB on top of MSA+DSG (ablation) | SHINE · DreamEditBench | ImageReward: 0.5745 -> 0.5709 | Hurt |
| use blurred Q_img for negative velocity instead of text negative prompts (ablation) | DSG on FLUX · FLUX robustness analysis | image quality | Improved |
| swap adapter for per-concept LoRA (ablation) | SHINE · DreamEditBench | DreamSim: 0.3730 -> 0.3577 | Improved |
| replace Flux.1 Fill with SHINE (training-free, FLUX) | image composition · ComplexCompo | DreamSim: 0.6018 -> 0.4294 (adapter); 0.3542 (LoRA) | Improved |
| apply SHINE to SDXL, SD3.5 and PixArt-Sigma instead of FLUX | SHINE · DreamEditBench | DINOv2: 0.7334 (SDXL), 0.7407 (SD3.5), 0.7445 (PixArt) vs 0.6745 Flux Fill | Improved |
Closest papers by meaning
- PS-Diffusion: Photorealistic Subject-Driven Image Editing with Disentangled Control and Attention (CVPR 2025 · similarity 0.87)
- RealCompo: Balancing Realism and Compositionality Improves Text-to-Image Diffusion Models (NeurIPS 2024 · similarity 0.87)
- MirrorVerse: Pushing Diffusion Models to Realistically Reflect the World (CVPR 2025 · similarity 0.86)
- $\textit{Bifr\"ost}$: 3D-Aware Image Compositing with Language Instructions (NeurIPS 2024 · similarity 0.86)
- Generating compositional scenes via Text-to-image RGBA Instance Generation (NeurIPS 2024 · similarity 0.86)
- CONFORM: Contrast is All You Need for High-Fidelity Text-to-Image Diffusion Models (CVPR 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.
