Research map / Image generation / Image restoration
CODE: Confident Ordinary Differential 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 CODE, a blind, training-free image editing method that inverts a pretrained diffusion model via the probability-flow ODE, applies Langevin score updates in latent space, and uses confidence-based clipping to restore out-of-distribution or corrupted images.
Problem
Conditioning diffusion-based image generation on noisy or out-of-distribution guidance images, where fidelity to the input and realism of the output trade off, without knowledge of the corruption type.
Approach
Replace SDEdit's Gaussian noise injection with deterministic probability-flow ODE inversion to a chosen latent depth, refine the latent with Langevin score-based updates, and clip latents to a confidence interval derived from the marginal distribution (CBC) before inversion; the decoupled noise level, step size, and latent depth give finer control over realism versus fidelity.
Key results
- Averaged over 47 corruptions on CelebA-HQ, CODE reduces FID from 47.84 (SDEdit) to 30.66, a 36% improvement.
- CODE attains PSNR-Input 19.61 vs 18.74 for SDEdit, about 5% higher fidelity to the input.
- Confidence-based clipping lowers average FID from 34.75 (CODE w/o CBC) to 30.65 with CBC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace SDEdit with ODE-based CODE | Diffusion editing on CelebA-HQ · CelebA-HQ, 47 corruptions | FID: 47.84 (42.29) -> 30.66 (16.21) | Helped |
| Replace SDEdit with ODE-based CODE | Diffusion editing on CelebA-HQ · CelebA-HQ, 47 corruptions | PSNR-Input: 18.74 (3.92) -> 19.61 (4.66) | Helped |
| Replace SDEdit with ODE-based CODE | Diffusion editing on CelebA-HQ · CelebA-HQ, 47 corruptions | SSIM-Input: 0.46 (0.21) -> 0.49 (0.22) | Helped |
| Replace SDEdit with ODE-based CODE | Diffusion editing on CelebA-HQ · CelebA-HQ, 47 corruptions | LPIPS-Source: 0.32 (0.13) -> 0.30 (0.12) | Helped |
| Add confidence-based clipping (ablation) | CODE · CelebA-HQ, 47 corruptions | FID: 34.75 (14.6) -> 30.65 (16.2) | Helped |
| Add confidence-based clipping (ablation) | CODE · CelebA-HQ, 47 corruptions | PSNR-Input: 19.4 (4.5) -> 19.6 (4.7) | Helped |
| Add confidence-based clipping (ablation) | CODE · CelebA-HQ, 47 corruptions | SSIM-Input: 0.48 (0.22) -> 0.49 (0.22) | Helped |
| Add confidence-based clipping (ablation) | CODE · CelebA-HQ, 47 corruptions | LPIPS-Source: 0.31 (0.11) -> 0.30 (0.12) | Helped |
| Increase Langevin step size (ablation) | CODE · Image editing | Realism and fidelity | Mixed |
| Optimize in deeper rather than shallow latent spaces (ablation) | CODE · Image editing across corruption types | Editing magnitude | Mixed |
| Increase the number of Langevin updates (ablation) | CODE · Image editing | Realism | Helped |
| Reduce the CBC confidence parameter (ablation) | CODE encoding · Corrupted images | Removal of unlikely pixels and noisy artifacts | Helped |
Surprising findings
- DDIM inversion alone with CBC lowers FID but sharply reduces fidelity, showing CBC alone is insufficient without Langevin refinement.
Limitations
- Trade-off between realism and fidelity requires a larger two-parameter grid search than SDEdit's single parameter.
- Hyperparameters such as latent depth and step size must be chosen per input, which the authors note is not one-size-fits-all.
- Evaluation on DDPM models for 256x256 face and LSUN datasets only.
Experimental setup
Datasets: CelebA-HQ LSUN-Bedroom LSUN-Church
Benchmarks: CelebA-HQ with 47 corruption types LSUN-Bedroom LSUN-Church
Metrics: FID PSNR SSIM LPIPS L2 distance
Baselines: SDEdit CodeFormer GFPGAN DiffBIR
Builds on: SDEdit DDIM DDPM Langevin dynamics score-based generative models
Closest papers by meaning
- PnP Inversion: Boosting Diffusion-based Editing with 3 Lines of Code (ICLR 2024 · similarity 0.88)
- Object-Aware Inversion and Reassembly for Image Editing (ICLR 2024 · similarity 0.86)
- High-Fidelity Diffusion-Based Image Editing (AAAI 2024 · similarity 0.86)
- CoDi: Conditional Diffusion Distillation for Higher-Fidelity and Faster Image Generation (CVPR 2024 · similarity 0.86)
- Inversion-Free Image Editing with Language-Guided Diffusion Models (CVPR 2024 · similarity 0.86)
- Schedule Your Edit: A Simple yet Effective Diffusion Noise Schedule for Image Editing (NeurIPS 2024 · 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.