Research map / Image generation / Editing and style transfer
Schedule Your Edit: A Simple yet Effective Diffusion Noise Schedule for 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 the Logistic Schedule, a logistic-function-based diffusion noise schedule that removes derivative singularities at t=0, improving DDIM inversion stability and content preservation for text-guided real-image editing without retraining.
Problem
DDIM inversion for real-image editing accumulates noise prediction errors, leading to poor content preservation and edit fidelity; the authors trace this partly to singularities in traditional (scaled linear and cosine) noise schedules at t=0.
Approach
Replace linear and cosine noise schedules with a logistic-shaped cumulative signal schedule (alpha_bar_t = 1/(1+exp(-k(t-t0)))) that gives a well-defined derivative at t=0, yielding steadier information perturbation and better inverted latents; it plugs into existing inversion-based editing methods without retraining.
Key results
- On attribute editing with Edit Friendly DDPM, the Logistic Schedule reduces structure distance from 35.66 to 17.37 (x10^-3) and raises PSNR from 20.70 to 24.78 versus the linear schedule.
- With Direct Inversion, the Logistic Schedule achieves structure distance 16.06 and PSNR 25.73, the best among the inversion variants tested.
- Across the four editing settings, the Logistic Schedule improves background preservation and structure metrics over linear and cosine schedules, while textual CLIP similarity is sometimes slightly lower than the linear schedule.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace the linear schedule with the Logistic Schedule | Text-guided editing with Edit Friendly DDPM · Attributes Editing | Structure Dist ×10^-3: 35.66 -> 17.37 | Helped |
| Replace the linear schedule with the Logistic Schedule | Text-guided editing with Zero-Shot Pix2Pix · Object Switch | PSNR: 19.93 -> 22.91 | Helped |
| Replace the linear schedule with the Logistic Schedule | Text-guided editing with StyleDiffusion · Style/Scene Transferring | CLIP textual similarity (%): 25.39 -> 24.77 | Hurt |
| Replace the linear schedule with the Logistic Schedule | Text-guided editing with MasaCtrl · Non-rigid Editing | Structure Dist ×10^-3: 30.83 -> 15.87 | Helped |
| Set k = 0.008 instead of k = 0.015 (ablation) | Logistic Schedule · Attributes Editing | Structure Dist ×10^-3: 17.37 -> 16.27 | Helped |
| Set k = 0.011 instead of k = 0.015 (ablation) | Logistic Schedule · Attributes Editing | Structure Dist ×10^-3: 17.37 -> 16.64 | Helped |
| Set k = 0.017 instead of k = 0.015 (ablation) | Logistic Schedule · Attributes Editing | Structure Dist ×10^-3: 17.37 -> 22.79 | Hurt |
| Set k = 0.029 instead of k = 0.015 (ablation) | Logistic Schedule · Attributes Editing | Structure Dist ×10^-3: 17.37 -> 27.82 | Hurt |
| Set t0 = int(0.4T) instead of int(0.6T) (ablation) | Logistic Schedule · Attributes Editing | Structure Dist ×10^-3: 17.37 -> 24.31 | Hurt |
| Set t0 = int(0.8T) instead of int(0.6T) (ablation) | Logistic Schedule · Attributes Editing | Structure Dist ×10^-3: 17.37 -> 29.47 | Hurt |
| Replace Stable Diffusion v1.5 with Stable Diffusion 2.1 (ablation) | Plug-and-Play editing with Logistic Schedule · Plug-and-Play editing | Structure Dist ×10^-3: 26.66 -> 34.74 | Hurt |
| Replace Stable Diffusion v1.5 with SDXL (ablation) | Plug-and-Play editing with Logistic Schedule · Plug-and-Play editing | Structure Dist ×10^-3: 26.66 -> 28.33 | Hurt |
Surprising findings
- Advanced SD variants (SD-2.1, SDXL) increase textual similarity but degrade content preservation relative to SD-1.5.
Limitations
- Textual CLIP alignment is sometimes slightly lower than the linear schedule.
- Hyperparameters k and t0 were chosen empirically and the singularity analysis relies on approximations (first-order Taylor expansion, continuous-time DDIM equivalence).
Experimental setup
Datasets: PIEBench FFHQ AFHQ MS-COCO
Benchmarks: PIEBench
Metrics: DINO-I structure distance PSNR LPIPS MSE SSIM CLIP visual similarity CLIP textual similarity
Baselines: linear schedule cosine schedule EF-DDPM pix2pix-zero StyleDiffusion MasaCtrl
Builds on: DDPM DDIM iDDPM Plug-and-Play Null-Text Inversion Direct Inversion
Closest papers by meaning
- Inversion-Free Image Editing with Language-Guided Diffusion Models (CVPR 2024 · similarity 0.91)
- An Edit Friendly DDPM Noise Space: Inversion and Manipulations (CVPR 2024 · similarity 0.91)
- Noise Map Guidance: Inversion with Spatial Context for Real Image Editing (ICLR 2024 · similarity 0.91)
- Editable Noise Map Inversion: Encoding Target-image into Noise For High-Fidelity Image Manipulation (ICML 2025 · similarity 0.90)
- High-Fidelity Diffusion-Based Image Editing (AAAI 2024 · similarity 0.90)
- DCI: Dual-Conditional Inversion for Boosting Diffusion-Based Image Editing (NeurIPS 2025 · similarity 0.89)
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.