Research map / Image generation / Editing and style transfer
DCI: Dual-Conditional Inversion for Boosting Diffusion-Based Image Editing
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 Dual-Conditional Inversion (DCI), which jointly conditions diffusion inversion on the source prompt and a VAE-encoded reference image via reference-guided noise correction plus fixed-point latent refinement, improving reconstruction and editing on PIE-Bench with Stable Diffusion v1.4.
Problem
Diffusion inversion methods trade off reconstruction accuracy against editing flexibility, because the inverted latent noise does not preserve both semantic alignment and structural consistency with the source image.
Approach
Treat inversion as a dual-conditioned fixed-point optimization: at each DDIM step, correct the prompt-conditioned noise prediction toward a reference noise derived from the source image latent, then refine the latent so it is a fixed point of the DDIM inversion update.
Key results
- On PIE-Bench with P2P editing, DCI reports DINO 6.07 (x10^-3) vs 8.81 for SPDInv, the second-best method, and PSNR 29.38 vs 28.60.
- Reconstruction MSE on 100 PIE-Bench captions: DCI 1.12e-4 vs SPDInv 1.21e-4 and DDIM 1.32e-4.
- Ablation shows K=5, lambda=2, eta=0.001 as the best default; lambda=5 and eta=0.01 degrade all metrics substantially.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply dual-conditional inversion instead of SPDInv | Stable Diffusion v1.4 with P2P editing · PIE-Bench | DINO ×10^3: 8.81 -> 6.07 | Helped |
| Apply dual-conditional inversion instead of SPDInv | Stable Diffusion v1.4 with P2P editing · PIE-Bench | PSNR: 28.60 -> 29.38 | Helped |
| Apply dual-conditional inversion instead of SPDInv | Stable Diffusion v1.4 with P2P editing · PIE-Bench | LPIPS ×10^3: 36.01 -> 33.01 | Helped |
| Apply dual-conditional inversion instead of SPDInv | Stable Diffusion v1.4 with P2P editing · PIE-Bench | MSE ×10^4: 24.54 -> 21.28 | Helped |
| Apply dual-conditional inversion instead of SPDInv | Stable Diffusion v1.4 with P2P editing · PIE-Bench | SSIM ×10^2: 86.23 -> 87.14 | Helped |
| Apply dual-conditional inversion instead of SPDInv | Stable Diffusion v1.4 with P2P editing · PIE-Bench | CLIP: 25.26 -> 25.52 | Helped |
| Use K=2 optimization rounds instead of K=5 (ablation) | DCI · PIE-Bench | PSNR: 29.38 -> 29.32 | No clear effect |
| Use K=10 optimization rounds instead of K=5 (ablation) | DCI · PIE-Bench | PSNR: 29.38 -> 29.29 | Hurt |
| Set reference-guided correction weight λ=1 instead of λ=2 (ablation) | DCI · PIE-Bench | PSNR: 29.38 -> 29.29 | Hurt |
| Set reference-guided correction weight λ=5 instead of λ=2 (ablation) | DCI · PIE-Bench | PSNR: 29.38 -> 28.25 | Hurt |
| Set fixed-point learning rate η=0.0001 instead of η=0.001 (ablation) | DCI · PIE-Bench | PSNR: 29.38 -> 28.80 | Hurt |
| Set fixed-point learning rate η=0.01 instead of η=0.001 (ablation) | DCI · PIE-Bench | PSNR: 29.38 -> 23.05 | Hurt |
Surprising findings
- Increasing the correction weight lambda to 5 or the refinement learning rate to 0.01 severely degrades performance, indicating sensitivity to these hyperparameters.
Limitations
- Gains over SPDInv on reconstruction CLIP score are marginal (26.94 vs 26.92)
- Performance is sensitive to lambda and eta hyperparameters
Experimental setup
Datasets: PIE-Bench COCO2017
Benchmarks: PIE-Bench
Metrics: DINO PSNR LPIPS MSE SSIM CLIP score
Baselines: DDIM inversion NTI NPI AIDI NMG DirectInv ProxEdit SPDInv h-Edit
Builds on: DDIM inversion Null-text inversion SPDInv Prompt-to-Prompt
Closest papers by meaning
- Inversion-Free Image Editing with Language-Guided Diffusion Models (CVPR 2024 · similarity 0.92)
- High-Fidelity Diffusion-Based Image Editing (AAAI 2024 · similarity 0.91)
- Tuning-Free Inversion-Enhanced Control for Consistent Image Editing (AAAI 2024 · similarity 0.91)
- Object-Aware Inversion and Reassembly for Image Editing (ICLR 2024 · similarity 0.90)
- PnP Inversion: Boosting Diffusion-based Editing with 3 Lines of Code (ICLR 2024 · similarity 0.90)
- Semantic Image Inversion and Editing using Rectified Stochastic Differential Equations (ICLR 2025 · similarity 0.90)
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.