The finished image is never one generation

The gap between a good image and a usable one

The image is right. The expression is right, the light is right, the client will like it. And there are six fingers on the left hand, the shop sign says CAFFEE, and there is a stray object floating near the shoulder.

Beginners regenerate. They roll the dice again, lose the good expression, get worse hands, and spend an afternoon and a month's credits chasing a single perfect roll. Working illustrators do not do this. They accept the frame as a base plate and repair it.

Inpainting is that repair, and it is where most of the real work in this craft happens.

How inpainting works, and the one setting that matters

You paint a mask over a region. The model regenerates only inside the mask, conditioned on everything outside it, so the new pixels match the surrounding light and colour.

Now the setting that separates people who can fix hands from people who cannot. Look for a control called "inpaint at full resolution", "only masked", or "crop to mask".

With it off, the model regenerates your masked region at the whole image's scale. If your picture is 1024 pixels wide and the hand occupies 80 of them, the model gets 80 pixels of budget to draw a hand. It cannot. It produces the same mush every time, and you conclude the model is bad at hands.

With it on, the tool crops to the masked area plus some padding, regenerates *that crop* at the model's full working resolution — so the hand gets 800 pixels of attention instead of 80 — and scales the result back in. The same model that could not draw a hand now draws it correctly, because the problem was never the model. It was the pixel budget.

That single toggle fixes more images than any prompt technique in existence.

The other three controls

Mask blur / feather. A hard mask edge leaves a visible seam. A few pixels of blur lets the new region fade into the old. Too much blur and the model starts blending the thing you were trying to remove back in.

Padding. How much surrounding context the crop includes. Too little and the model cannot see the arm the hand belongs to, so it draws a hand pointing the wrong way. Too much and you are back to a small pixel budget.

Denoise strength, again. For a repair, 0.6-0.8 — you want it to actually replace what is there. For a subtle correction, 0.3.

Outpainting, and the thing clients always ask for

Outpainting extends the canvas: the model invents new image beyond the existing edges, matched to what is there.

The everyday use is not artistic. You made a beautiful 16:9 frame and the client needs 9:16 for Stories and 1:1 for the feed. Cropping destroys the composition. Outpainting builds the missing top and bottom, and then you crop from a larger canvas.

Photoshop calls this Generative Expand. Free equivalents that do the same job: Krita with AI Diffusion, InvokeAI's canvas, ComfyUI with an outpaint node, and most hosted tools have an expand or uncrop button.

The failure modes, with their causes

Colour drift. Inpaint the same area eight times and the region slowly diverges in colour and contrast from the rest, because each pass is a fresh generation matched only approximately to its surroundings. Fix it outside the model: flatten, then a curves and colour-balance pass on that region in a raster editor. Do not try to inpaint your way back to matching.

Halo edges. Almost always too little padding or a mask drawn tight to the object's silhouette. Mask generously — include some of the background around what you are replacing.

The model ignores your mask. Some tools mask the *latent* rather than the pixels, which means the edges bleed. If a tool consistently modifies outside your mask, it is that, and switching tools is faster than fighting it.

The workflow this implies

1. 1Generate cheaply until you have a base plate with the right composition and mood. Stop there. Do not chase perfection.
2. 2Inpaint each defect separately, at full resolution, one at a time.
3. 3Flatten and do a colour and contrast pass in an editor.
4. 4Only then upscale.

That order matters and lesson seven explains why the last step is last.

Today: take any generated image you have with a flaw in it, open a free inpainting tool, mask the flaw, and confirm you have found the "only masked" setting before you press generate. If the tool does not have one, that tells you something about the tool.