LEDITS¶

Development line: project:ledits · thread ledits-development
Last event: 2023-12-03 · 2 dated since 2023-07-04 · Researched: 2026-09-04 · confidence: medium

What it is¶

LEDITS is an image editing pipeline for modifying real images with semantic text prompts instead of a generic img2img pass.

• Inversion: inverts the input image before editing.
• Guidance: applies one or several concept prompts.
• Masking: uses implicit masks to confine changes to relevant regions.
• Integration: exposes Stable Diffusion and SDXL pipelines in Diffusers.

The original LEDITS Space is paused, and the Diffusers implementation does not guarantee perfect inversion. Use LEDITS++ locally for Stable Diffusion or SDXL semantic edits; use the authors’ implementation when inversion fidelity is the deliverable.

Development line¶

• 2023-07-04 — LEDITS public demo and Hugging Face Space were linked. On 2023-07-04, the LEDITS development thread linked a public web demo, the Hugging Face Space, and the repository for editing-images/ledits. These links do not establish release scope, model version, or capabilities.
• 2023-12-03 — LEDITS++ public demo and Hugging Face Space were linked. On 2023-12-03, the thread linked a public LEDITS++ demo, the Hugging Face Space, and the repository for editing-images/leditsplusplus. It also linked back to LEDITS. These links do not confirm the exact technical relationship between LEDITS and LEDITS++.

What changed¶

• 2023-07-04 — LEDITS combined Edit-Friendly DDPM inversion with SEGA semantic guidance for real-image editing on Stable Diffusion v1.5.
• 2023-12-03 — LEDITS++ introduced SDE-DPM-Solver++ inversion, per-edit implicit masking, and native simultaneous edits.

How to use this¶

As of 2023-12-03, distinguish the original LEDITS public artifacts from the separately linked LEDITS++ artifacts and consult the matching Hugging Face Space or repository for the variant you intend to use.

1. Choose LEditsPPPipelineStableDiffusion for Stable Diffusion or LEditsPPPipelineStableDiffusionXL for SDXL, then load a compatible base checkpoint on CUDA. — https://huggingface.co/docs/diffusers/api/pipelines/ledits_pp
2. Call invert(image=...) for each new source image before invoking the editing pipeline; the edit always uses the last inversion. — https://huggingface.co/docs/diffusers/api/pipelines/ledits_pp
3. Use source_prompt only when guided inversion helps; an empty source prompt disables that guidance. — https://huggingface.co/docs/diffusers/api/pipelines/ledits_pp
4. Pass one or more editing_prompt values and align reverse_editing_direction, edit_guidance_scale, and edit_threshold with those prompts; direction controls whether each concept is increased or decreased. — https://huggingface.co/docs/diffusers/api/pipelines/ledits_pp
5. Inspect the inversion reconstruction and edit result; if exact inversion is required, use the authors’ repository rather than the Diffusers implementation. — https://github.com/ml-research/ledits_pp

Best practices¶

• Use a seeded torch.Generator when comparing parameter changes so inversion is deterministic. — https://huggingface.co/docs/diffusers/api/pipelines/ledits_pp
• Treat skip as an edit-strength control: lower values make stronger changes to the input image. — https://huggingface.co/docs/diffusers/api/pipelines/ledits_pp
• Set edit_threshold to the edit footprint: use a smaller value for whole-image changes such as style transfer, and a value proportional to the affected region for local edits. — https://arxiv.org/html/2311.16711v1
• Keep the implicit intersection mask by default; add user_mask only when the automatic mask does not match the required region. — https://huggingface.co/docs/diffusers/api/pipelines/ledits_pp
• For research reproduction or a hard reconstruction-fidelity requirement, prefer the authors’ implementation because its README states that Diffusers does not guarantee perfect inversion. — https://github.com/ml-research/ledits_pp

Superseded by this¶

• 2023-07-04 — Treating the original LEDITS Gradio Space as an operational entry point is obsolete: the Space is paused; retain it as historical code rather than a dependable workflow.
• 2023-12-03 — Hosted-demo-only usage is obsolete: since 2024-03-15, Diffusers has shipped LEDITS++ pipeline APIs for Stable Diffusion and SDXL.
• 2023-12-03 — Treating the paper’s perfect-reconstruction result as a blanket library guarantee is obsolete: the authors state that the Diffusers implementation does not guarantee perfect inversion.

Still unknown¶

• The primary sources do not explicitly label LEDITS++ as a versioned replacement for the 2023 LEDITS demo; they are treated here as closely related work, not a proven product-version lineage.
• The LEDITS++ Space is displayed as Running on Zero, but no image generation was executed during this check; that is not proof of dependable hosted availability.
• No current cross-model evaluation against newer general-purpose image editors was identified, so this does not claim that LEDITS++ remains best-in-class outside its Stable Diffusion and SDXL integration.

Sources¶

Agent brief¶

• Subject: project:ledits, thread ledits-development, 2 dated events 2023-07-04 → 2023-12-03.
• Practical note: As of 2023-12-03, practitioners should distinguish the original LEDITS public artifacts from the separately linked LEDITS++ artifacts and consult the corresponding Hugging Face Space or repository for the variant they intend to use.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.