Research map / Image generation / Text-to-image diffusion
LEDITS++: Limitless Image Editing using Text-to-Image Models
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.
Introduces LEDITS++, a tuning-free real-image editing method for diffusion models that combines a perfect DPM-Solver++-based inversion with multi-concept guidance and implicit attention/noise masking, plus the TEdBench++ benchmark.
Problem
Real image editing with text-to-image diffusion models is slow (fine-tuning or optimization), deviates strongly from the input image, and does not support multiple simultaneous edits.
Approach
Derives a perfect-reconstruction inversion for the second-order SDE DPM-Solver++ sampler, applies independent per-concept semantic guidance terms, and grounds each edit with an implicit mask that intersects cross-attention maps and noise-estimate difference maps.
Key results
- LEDITS++ with SD1.5 reaches 0.79 success rate on TEdBench++ with LPIPS 0.30, versus Imagic w/ SD1.5 at 0.58 and LPIPS 0.57.
- Perfect reconstruction (RMSE 0.00) at 1.78 s execution time, versus 37.23 s for vanilla DDIM inversion and 349.98 s for Imagic.
- On CelebA multi-attribute editing, LEDITS++ is closest to the ideal CLIP-vs-LPIPS trade-off, with only LEDITS++ and edit-friendly DDPM reliably reaching CLIP above 0.25.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LEDITS++ instead of Imagic | Stable Diffusion 1.5 image editing · TEdBench | success rate (SR): 0.55 -> 0.75 | Helped |
| Use LEDITS++ instead of Imagic | Stable Diffusion 1.5 image editing · TEdBench | LPIPS: 0.56 -> 0.28 | Helped |
| Use LEDITS++ instead of Imagic | Stable Diffusion 1.5 image editing · TEdBench++ | success rate (SR): 0.58 -> 0.79 | Helped |
| Use LEDITS++ instead of Imagic | Stable Diffusion 1.5 image editing · TEdBench++ | LPIPS: 0.57 -> 0.30 | Helped |
| Use LEDITS++ instead of Imagic | TEdBench · TEdBench | success rate (SR): 0.83 -> 0.84 | Helped |
| Use LEDITS++ instead of Imagic | TEdBench · TEdBench | LPIPS: 0.59 -> 0.33 | Helped |
| Use LEDITS++ instead of Edit-friendly DDPM | Stable Diffusion 1.5 image editing · Image editing | execution time (s): 10.36 ±0.05 -> 1.78 ±0.03 | Helped |
| Use LEDITS++ instead of Vanilla DDIM Inversion | Stable Diffusion 1.5 image editing · Image editing | execution time (s): 37.23 ±0.04 -> 1.78 ±0.03 | Helped |
Surprising findings
- Imagic often generated a completely new image disregarding the input, despite being a fine-tuning method built for preservation.
Limitations
- Edit quality depends on the capability of the underlying pre-trained diffusion model; concepts poorly represented in the model may fail.
- Generic prompts can change object identity, especially at strong guidance hyperparameters.
- Implicit masks cannot always infer user intent, so user masks are sometimes needed.
Experimental setup
Datasets: CelebA COCO panoptic segmentation TEdBench
Benchmarks: CelebA multi-attribute editing TEdBench TEdBench++ COCO panoptic segmentation
Metrics: CLIP score LPIPS success rate reconstruction RMSE execution time IoU
Baselines: SDEdit Imagic DDIM Inversion Pix2Pix-Zero DiffEdit Edit-friendly DDPM
Builds on: DDPM inversion (Edit-friendly DDPM) DPM-Solver++ classifier-free guidance Prompt-to-Prompt SEGA-style semantic guidance
Closest papers by meaning
- Inversion-Free Image Editing with Language-Guided Diffusion Models (CVPR 2024 · similarity 0.89)
- Towards Efficient Diffusion-Based Image Editing with Instant Attention Masks (AAAI 2024 · similarity 0.89)
- BARET: Balanced Attention Based Real Image Editing Driven by Target-Text Inversion (AAAI 2024 · similarity 0.89)
- Object-Aware Inversion and Reassembly for Image Editing (ICLR 2024 · similarity 0.88)
- EditInfinity: Image Editing with Binary-Quantized Generative Models (NeurIPS 2025 · similarity 0.88)
- SwiftEdit: Lightning Fast Text-Guided Image Editing via One-Step Diffusion (CVPR 2025 · similarity 0.88)
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.