Research map / Image generation / Editing and style transfer
FDS: Frequency-Aware Denoising Score for Text-Guided Latent Diffusion 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 FDS, which applies wavelet decomposition to latents in score distillation (DDS) and uses stop-gradient on selected frequency subbands to edit images and 3D textures while preserving detail and color.
Problem
Text-guided image editing with T2I diffusion via score distillation often loses high-frequency local detail and introduces unintended color changes because gradients update all frequency bands indiscriminately.
Approach
Decompose the latent with a discrete wavelet transform into low- and high-frequency subbands and apply a stop-gradient to either group during DDS optimization, so only the desired frequency bands are updated; extended to frequency-decomposed triplanes for 3D texture editing.
Key results
- On 200 LAION-COCO cat-edit images, the method reaches CLIP 23.65 (matching CDS), LPIPS 0.129 (vs CDS 0.136, DDS 0.297), and SSIM 0.819 (vs CDS 0.818, DDS 0.736).
- In a 24-user study, over 80% on average preferred the method's edits over CDS (86.3%) and DDS (82.1%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add wavelet-based selective frequency optimization (ablation) | DDS · 2D image editing (200 LAION-COCO cat images) | LPIPS: 0.297 -> 0.129 | Helped |
| Add wavelet-based selective frequency optimization (ablation) | DDS · 2D image editing (200 LAION-COCO cat images) | SSIM: 0.736 -> 0.819 | Helped |
| Add wavelet-based selective frequency optimization (ablation) | DDS · 2D image editing (200 LAION-COCO cat images) | CLIP score: 23.64 -> 23.65 | No clear effect |
| Optimize low-frequency components to preserve high-frequency details | FDS · Text-guided image editing | user preference: 90.5% | Helped |
| Optimize high-frequency components to preserve color fidelity | FDS · Text-guided image editing | user preference: 92.9% | Helped |
| Use Haar wavelets (DB=1) rather than smoother Daubechies wavelets (ablation) | FDS · Coffee-to-matcha latte art editing | localization | Hurt |
| Increase decomposition level to J=5 (ablation) | FDS · Coffee-to-matcha latte art editing | localization | Hurt |
| Use DFT or DCT instead of wavelet decomposition (ablation) | FDS · Dog-to-cat drawing edit | spatial localization | Hurt |
Limitations
- Inherits limitations of diffusion models and score distillation, including difficulty interpreting prompts.
- Requires a lengthy sampling/optimization process (N=500 iterations).
- Larger decomposition levels reduce localization, causing color spill-over into adjacent regions.
Experimental setup
Datasets: LAION-COCO
Benchmarks: LAION-COCO subset of 200 cat images
Metrics: CLIP score LPIPS SSIM user preference
Baselines: DDS CDS SDS NMG Direct Inversion PnP Prompt-to-Prompt DiffuseIT FlexiEdit FreeDiff DreamSampler
Builds on: Delta Denoising Score Score Distillation Sampling Contrastive Distillation Score discrete wavelet transform
Closest papers by meaning
- Disentangling Dual Image References in Frequency Aware Diffusion Models for Personalized Generation (NeurIPS 2026 · similarity 0.88)
- One Stone with Two Birds: A Null-Text-Null Frequency-Aware Diffusion Models for Text-Guided Image Inpainting (NeurIPS 2025 · similarity 0.88)
- Frequency-Controlled Diffusion Model for Versatile Text-Guided Image-to-Image Translation (AAAI 2024 · similarity 0.88)
- Contrastive Denoising Score for Text-guided Latent Diffusion Image Editing (CVPR 2024 · similarity 0.87)
- TiNO-Edit: Timestep and Noise Optimization for Robust Diffusion-Based Image Editing (CVPR 2024 · similarity 0.86)
- FreqTS: Frequency-Aware Token Selection for Accelerating Diffusion Models (AAAI 2025 · 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.