Inpainting
Filling in a masked region so it matches the rest — commercially the most useful generative feature, and the one that quietly ended photographic evidence.
When not to use it
- When the result will be treated as evidence. The fill is plausible, not true, and nothing marks the difference.
- For precise insertion of a specific object. Lighting, perspective and shadow are where it fails, subtly.
- On large regions of a complex scene. Not enough context reaches the middle, and it invents.
- When a clone-stamp would do. For small, simple removals, deterministic tools are faster and don't hallucinate.
Reach for something else instead
- Clone stamp / content-aware fill — deterministic, predictable, fine for small holes.
- Reshooting — if the object shouldn't be in the frame, sometimes moving the camera is the answer.
- Compositing — for insertion, a real cut-out with manual lighting beats a generated one.
- Classical inpainting — for scratches and dust, where you want interpolation rather than invention.
Sources & further reading
- Bertalmío et al. (2000), Image Inpainting — the classical formulation; structure propagation before there was content generation.
- Suvorov et al. (2021), Resolution-robust Large Mask Inpainting with Fourier Convolutions — LaMa; why a global receptive field matters for large holes.
- Lugmayr et al. (2022), RePaint: Inpainting using Denoising Diffusion Probabilistic Models — inpainting from an unmodified pretrained model, at inference time.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Hard mask edges. Feather them, or the model has a discontinuity it can't reconcile and you get a halo.
- Not prompting the fill. Without a hint the model guesses from texture alone.
- Accepting the first result. It's stochastic — the workflow is generate-and-pick.
- Treating a filled region as recovered rather than invented. There is no mechanism for truth here.
At a glance
FieldGenerative AI
Removalnearly solved, enormously used
Insertionhard; lighting, perspective, shadow
Mechanismkeep known region fixed at each denoising step
Fails atmask edges
Consequenceremoval leaves no forensic artefact
DifficultyBeginner
Flashcards for this concept
Question
—
Answer
—
1 / 4
Often compared with
Inpainting vs. text-to-image — one fixes a photo that exists, the other invents one. The first is used far more, and generates most of the revenue.