What inpainting actually does

Inpainting is the process of regenerating a selected region of an existing image using a diffusion model, conditioned on both the surrounding pixels and your text prompt. Unlike full image generation, inpainting operates within tight visual constraints — the model must produce a result that blends seamlessly with the unchanged areas of the image, matching lighting, colour temperature, texture scale, and perspective. This makes it a fundamentally different task from generation from scratch, and one that requires different techniques to execute well.

For face work specifically, inpainting is the primary tool for correcting artefacts in generated or face-swapped outputs — fixing a slightly malformed ear, correcting an unnatural eye, adjusting skin tone in a blending boundary zone. It is also the technique that makes composite face work possible at a quality level that full generation rarely achieves, because it allows you to anchor the face in a real photographic context rather than generating everything from noise.

Mask construction: the skill that determines everything

The quality of an inpainting result is determined primarily by the quality of the mask — the selection that defines which pixels the model regenerates and which it preserves. Masks that are too small force the model to work in a cramped region, producing outputs that look correct in isolation but fail to integrate with the surrounding image. Masks that are too large give the model too much freedom and destabilise the regions you intended to preserve.

For face correction work, the mask should extend several pixels beyond the visible problem area in every direction, typically 10 to 20 pixels of feathering depending on the resolution of the working image. This feathering zone gives the model room to blend the regenerated region into the surrounding pixels rather than creating a hard edge. Hard-edged masks are the single most common cause of visible seams in inpainting work, and they are almost entirely avoidable with proper mask construction.

Prompt strategies for inpainting vs. generation

Prompting for inpainting requires a different approach from prompting for full generation. Because the model is conditioned on the surrounding image, many aspects of the output are already constrained by context — lighting direction, colour temperature, skin tone, depth of field. Your prompt should describe the local region you are regenerating, not the entire image. For a face correction task, a prompt like 'natural skin texture, photorealistic, soft light' is more effective than a full portrait description, because the model uses the image context to fill in the details your prompt does not specify.

Over-specifying in the inpainting prompt is a common error. If the surrounding image already shows warm afternoon light and your prompt specifies 'cool studio lighting', the model must choose between prompt and context — and the resulting blend typically satisfies neither. Describe what you want in the masked region, not what you would want in a full generation from scratch, and let the surrounding image do the work of establishing the broader visual context.

Denoising strength and how to calibrate it

Denoising strength is the inpainting parameter with the highest impact on the balance between regeneration and preservation. At a value of 1.0, the model treats the masked region as pure noise and generates it entirely from scratch, constrained only by the surrounding pixels and your prompt. At a value of 0.3, it makes only subtle adjustments to the existing content in the masked region, preserving most of the original detail while introducing small corrections.

For most face correction tasks — fixing an artefact, smoothing a boundary, adjusting an expression — a denoising strength of 0.5 to 0.7 is the productive range. Below 0.5, changes are often too subtle to fully resolve the problem. Above 0.75, the model begins to deviate significantly from the original content and identity drift becomes visible. The exception is full face replacement or heavy compositing work, where higher denoising strength — 0.85 to 1.0 — is appropriate because you want the model to generate freely within the masked region rather than preserve the original.

Multi-pass inpainting for complex corrections

Complex face corrections rarely succeed in a single inpainting pass. Professional workflows use multiple passes at progressively lower denoising strengths, each targeting a more refined aspect of the correction. The first pass, at higher denoising strength, establishes the broad structural correction — fixing the underlying geometry of the problem area. Subsequent passes, at lower denoising strengths, refine texture, and correct any new artefacts introduced by the first pass.

This multi-pass approach is particularly effective for face-swap boundary work, where the transition between the inserted face and the surrounding skin or hair is rarely perfect in a single pass. A first pass at 0.7 denoising strength corrects the broad lighting mismatch; a second pass at 0.4 across the boundary zone smooths the texture transition; a final pass at 0.25 with a large feathered mask over the entire face unifies the colour grade and eliminates any remaining visible seam. The total time investment is higher than a single pass.