Inpaint Image (LaMa)

How it fits

Typical backends

Caveats

• Hard constraintThe mask must match the source's width and height exactly. A mask with no set pixels is a no-op and the source is returned unchanged; an empty (zero-size) source Image likewise passes through. Every input emits exactly one image.
• I/O contractThe hole specification is a Mask carrier (any set pixel is a hole) or a second Image channel reduced to grayscale (any non-zero pixel is a hole). Dilate the mask a few pixels past the object so its outline is fully covered.
• I/O contractLaMa pads the image and mask to a multiple of eight internally and crops the result back; the output always matches the input resolution exactly.
• Parameter interactionmodel `lama` ignores prompt, steps, guidance, and seed; it is deterministic. model `sd2` reads all four to steer the latent-diffusion fill.
• Parameter interactionUnder model `sd2`, fix seed to a non-negative value for a reproducible fill; the diffusion backbone produces a different result per call otherwise.
• I/O contractThe output preserves the source's original colour space.
• CostBoth backbones load their weights from the bound model bundle with no runtime download — the big-lama TorchScript checkpoint (~196 MB) for LaMa or the multi-gigabyte Stable Diffusion 2 inpainting snapshot for sd2. Allow a generous startup budget on a cold deployment.
• Fallbackdevice `cuda` falls back to CPU with a printed warning when no GPU is visible. LaMa stays usable on CPU; model `sd2` becomes impractically slow.
• CompatibilityLaMa weights are Apache-2.0; Stable Diffusion 2 inpainting is CreativeML Open RAIL++-M — review it before deploying model `sd2` commercially.