Diffusers guide: inpainting (with image-to-image chaining)
Source: https://huggingface.co/docs/diffusers/using-diffusers/inpaint
Diffusers library repo itself is already in the seed catalog (huggingface--diffusers); this record covers the inpaint/img2img teaching content only.
Structured fields
{
"checked_at": "2026-10-17T00:00:00.000Z",
"usefulness": "Foundational, runnable-code walkthrough of mask-based inpainting in Diffusers: mask conventions, strength, guidance_scale, padding_mask_crop, mask blur, apply_overlay to keep unmasked pixels, chaining text-to-image -> inpaint -> image-to-image, and ControlNet inpaint conditioning. The sibling img2img guide covers strength and chaining for whole-image edits.",
"source_type": "documentation",
"availability": "available",
"dependencies": [
"diffusers",
"torch"
],
"source_owner": "Hugging Face",
"next_check_at": "2026-10-24T00:00:00.000Z",
"hardware_notes": "Guide recommends enable_model_cpu_offload and PyTorch 2 SDPA or xFormers to reduce memory; no VRAM figures given.",
"known_limitations": [
"Examples target SD 1.5 / SDXL / Kandinsky 2.2 class models; instruction-edit and flow-transformer models (FLUX, Qwen-Image) have different editing interfaces.",
"Docs snapshot is versioned (v0.41.0 at check time); parameters may change."
]
}Evidence claims
- [sourced_fact] White mask pixels are filled from the prompt and black pixels are kept; page documents strength, guidance_scale, negative_prompt and padding_mask_crop parameters. source
- [sourced_fact] Inpaint-specific checkpoints give more natural transitions but are more likely to change unmasked areas; VaeImageProcessor.apply_overlay() can force unmasked pixels to remain unchanged at the cost of less natural transitions. source
- [sourced_fact] Image-to-image guide: strength sets how much noise is added (1.0 means the initial image is more or less ignored) and relates to num_inference_steps; chaining pipelines with output_type=latent only works when pipelines share a VAE. source
- [inference] The pages render docs for diffusers v0.41.0 at the check date; examples use SD 1.5, SDXL and Kandinsky 2.2, not newer DiT/flow models. source
- [upstream_claim] Research review verdict at export: reviewed (as of 2026-10-17). Recorded by the research pipeline, not a hub review outcome; evidence.review stays unreviewed until an independent reviewer records one.
Rights
| component | license | hosting | redistribution | modification | training | Evidence and notes |
|---|---|---|---|---|---|---|
| documentation | unknown | unknown | unknown | unknown | unknown | Docs-specific terms not checked. The huggingface/diffusers repository LICENSE is Apache-2.0 (fetched https://github.com/huggingface/diffusers/blob/main/LICENSE) but that was not confirmed to cover the rendered documentation site. |
| code | unknown | unknown | unknown | unknown | unknown | Example code snippets reference third-party checkpoints (SD 1.5, SDXL, Kandinsky, community style models) with their own licenses; not evaluated here. |