Research map / Image generation / Editing and style transfer
DreamFuse: Adaptive Image Fusion with Diffusion Transformer
Junjia Huang, Pengxiang Yan, Jiyang Liu, Jie Wu, Zhao Wang, Yitong Wang, Liang Lin, Guanbin Li
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 |
|---|---|---|---|
| use positional affine instead of direct transform for foreground position (ablation) | DreamFuse first stage · DreamFuse test set | VisionReward: 4.72 -> 5.28 | Improved |
| use positional affine instead of mask tokenizer (ablation) | DreamFuse first stage · DreamFuse test set | VisionReward: 4.93 -> 5.28 | Improved |
| use positional affine instead of direct transform (ablation) | DreamFuse first stage · DreamFuse test set | CLIP: 35.04 -> 35.22 | Improved |
| add LDPO second-stage preference optimization on seed pairs (ablation) | DreamFuse · DreamFuse test set | ImageReward: 1.07 -> 1.16 | Improved |
| add LDPO on seed pairs (ablation) | DreamFuse · DreamFuse test set | VisionReward: 5.28 -> 5.52 | Improved |
| add copy-paste negative loss L_LDPO (ablation) | DreamFuse w/ LDPO · DreamFuse test set | CLIP: 35.36 -> 35.51 | Improved |
| add L_noise loss (ablation) | DreamFuse w/ LDPO + L_LDPO · DreamFuse test set | VisionReward: 5.58 -> 5.56 | No clear effect |
| lower text dropout probability p to 80% (ablation) | DreamFuse · DreamFuse test set | CLIP: p=99% chosen over 80% | Hurt |
| set dilation factor alpha to 0 (copy-paste treated as preferred) (ablation) | DreamFuse LDPO · DreamFuse test set | VisionReward: alpha 0 gives significant VR decline vs alpha 1.5 | Hurt |
| set dilation factor alpha to 1 (only bounding box region) (ablation) | DreamFuse LDPO · DreamFuse test set | VisionReward: alpha 1 gives a performance drop vs alpha 1.5 | Hurt |
| DreamFuse versus second-best method on TF-ICON dataset | DreamFuse · TF-ICON dataset | VisionReward: 2.37 -> 3.27 | Improved |
| DreamFuse versus second-best method on DreamFuse test set | DreamFuse · DreamFuse test set | VisionReward: 4.14 -> 5.56 | Improved |
Closest papers by meaning
- DreamFuse: Toward Realistic and Seamless Image Fusion Across Diverse Scenarios (TPAMI 2026 · similarity 0.97)
- Text-DiFuse: An Interactive Multi-Modal Image Fusion Framework based on Text-modulated Diffusion Model (NeurIPS 2024 · similarity 0.86)
- GeminiFusion: Efficient Pixel-wise Multimodal Fusion for Vision Transformer (ICML 2024 · similarity 0.86)
- Revisiting Generative Infrared and Visible Image Fusion Based on Human Cognitive Laws (NeurIPS 2025 · similarity 0.85)
- Efficient Rectified Flow for Image Fusion (NeurIPS 2025 · similarity 0.85)
- DreamLight: Towards Harmonious and Consistent Image Relighting (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.
