Research map / Image generation / Editing and style transfer

h-Edit: Effective and Flexible Diffusion-Based Editing via Doob's h-Transform

CVPR 2025 · Image generation › Editing and style transfer · approach: Inversion-based diffusion editing · 4 citations (OpenAlex)

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 h-Edit, a training-free diffusion image editing method framed as reverse-time bridge modeling via Doob's h-transform and Langevin Monte Carlo, splitting each update into reconstruction and editing terms.

Problem

Diffusion-based image editing balances effective edits against faithful preservation of unedited content, and most training-free methods lack a theoretical foundation and struggle to combine multiple editing types.

Approach

Modify the reverse diffusion transition with Doob's h-transform so the backward process converges to the distribution p(x0)h(x0,0); sample edits with Langevin Monte Carlo, decomposing each update into a reconstruction term (from inversion) plus an editing gradient of log h, allowing product-of-experts composition of multiple editing objectives.

Key results

• On PIE-Bench, h-Edit-D + P2P improves over NT in LPIPS by 1.22e-2 and local CLIP by 0.017 (deterministic inversion setting).
• In face swapping on CelebA-HQ, h-Edit-R (implicit) reaches ID similarity 0.80 vs 0.74 for EF and 0.70 for FaceShifter, with LPIPS 0.04.
• h-Edit-R outperforms EF and LEDITS++ on all PIE-Bench metrics with or without P2P.

Measured effects reported in the paper

Surprising findings

• Some baselines (PnP Inv, NMG) achieve high faithfulness by largely reconstructing the original image without making meaningful edits in challenging cases.

Limitations

• Some difficult editing cases remain unsolved; the method sometimes fails to make the desired edit.
• Multiple optimization steps improve effectiveness but add cost; hyperparameters (guidance weights) may need manual tuning.
• Editing term modifies the whole image, which can reduce faithfulness in non-edited regions.

Experimental setup

Datasets: PIE-Bench CelebA-HQ

Benchmarks: PIE-Bench CelebA-HQ face swapping

Metrics: CLIP similarity local CLIP similarity PSNR LPIPS SSIM DINO distance ArcFace identity similarity expression error pose error FID

Baselines: Null-text inversion Negative prompt inversion StyleDiffusion Noise Map Guidance PnP Inversion Edit Friendly LEDITS++ FaceShifter MegaFS AFS DiffFace

Builds on: Edit Friendly DDIM Prompt-to-Prompt classifier-free guidance Langevin Monte Carlo

Closest papers by meaning

• High-Fidelity Diffusion-Based Image Editing (AAAI 2024 · similarity 0.91)
• PnP Inversion: Boosting Diffusion-based Editing with 3 Lines of Code (ICLR 2024 · similarity 0.90)
• Inversion-Free Image Editing with Language-Guided Diffusion Models (CVPR 2024 · similarity 0.90)
• FunEditor: Achieving Complex Image Edits via Function Aggregation with Diffusion Models (AAAI 2025 · similarity 0.89)
• Schedule Your Edit: A Simple yet Effective Diffusion Noise Schedule for Image Editing (NeurIPS 2024 · similarity 0.88)
• DCI: Dual-Conditional Inversion for Boosting Diffusion-Based Image Editing (NeurIPS 2025 · similarity 0.88)