Inpaint in Forge without repainting the whole image

Open Img2img and choose the Inpaint tab. Current source also exposes Inpaint sketch, Inpaint upload and Batch as separate modes.

Ordinary img2img lets the whole source influence the transformation. Inpainting adds a mask so the edit contract targets the painted region or its inverse.

Inpainting changes pixels inside the existing canvas. Outpainting enlarges the canvas and generates beyond an original edge.

Use Inpaint masked when the painted area should change. Use Inpaint not masked only when the painted area should be protected and everything outside it should change.

In the current Inpaint upload path, the mask is thresholded and the white region is treated as the editable region before an optional Mask mode inversion.

It keeps the source as the starting visual signal inside the mask. It is the clearest first choice when the old structure is useful.

Forge fills the region from surrounding image color before diffusion. It can help removal, but it is not a semantic object eraser and may produce flat or grey starts in some reported setups.

Use it as a comparison when the source structure strongly resists removal, addition or replacement. It gives diffusion a noisy starting latent inside the mask, so it usually needs more control from prompt and context.

The current processing path zeros the masked part of the initial latent. It is another starting condition, not a stronger version of the prompt and not a guaranteed repair.

Use Whole picture when global scene context or a broad edit matters. Use Only masked when a small region needs more processing resolution; supply enough padding for scale, lighting and attachment context.

The cropped mask region is resized to the processing width and height before being pasted back, so more of the diffusion budget is concentrated on that local area.

It expands the crop around the mask. Padding is context, not extra editable area: too little can lose scale; too much reduces local concentration.

There is no universal number. The inspected UI starts at 4. Compare small changes around that value and inspect the edge; a very large blur can overwhelm a tiny mask.

No one value fits repair, removal, replacement and addition. Compare controlled low, middle and high probes with the same mask and seed, then narrow the interval that produces enough change without scene drift.

Check for a blank or tiny effective mask, reversed Mask mode, Soft inpainting preservation, a prompt that still describes the old object, or a model/setup that also fails outside inpainting.

Check Mask mode, Whole picture processing, blur and overlay behavior. Some blending near the boundary is expected; broad unrelated change is a reason to rebuild a clean baseline and compare the saved source.

Mask both the object and shadow, remove the object from the prompt, describe the revealed background, and compare Fill against Latent noise one variable at a time.

Mask a plausible footprint, describe placement and contact, compare Latent noise, then send the rough successful form back and refine it with a smaller mask and Original.

Mask the complete malformed hand plus wrist context, describe the intended gesture and view, use Only masked with enough padding for the arm direction, and expect several controlled iterations.

Mask the face and choose Inpaint not masked so the painted face is protected. Inspect the boundary carefully because hair, neck, lighting and pose cross that protection edge.

Yes. Choose Inpaint upload and provide the source image plus a mask. Keep both images the same dimensions and confirm the intended white region before generating.

The current batch path supports an inpaint mask directory. One mask can be reused for all images; otherwise Forge matches masks by the input filename stem and skips inputs without a match. Prove one image and mask before a folder.

The core workflow can run with ordinary compatible checkpoints, as dated tutorials demonstrate. A dedicated inpainting checkpoint may behave differently; follow its publisher and base-family requirements instead of assuming a filename makes it compatible.

A dated video demonstrates one FLUX NF4 workflow, but open original-Forge issues report inconsistent behavior for Fill, Only masked, Hires and FLUX Fill models. Support for an arbitrary FLUX combination is UNKNOWN.

The bundled source describes mask-opacity-aware blending of original and generated content. At the inspected commit the source file exists, but an open issue reports that its UI is missing, and other reports describe inconsistent results. Treat it as optional and version-specific.

Save the source and mask; record original Forge commit, checkpoint, VAE, prompt, seed, dimensions, sampler, denoising, mask mode, masked content, inpaint area, blur, padding, extensions and the exact result selected for the next pass.