One broken hand should not cost you a great image. It costs you about ninety seconds.
One broken hand should not cost you a great image. It costs you about ninety seconds.
You finally got the image. The pose, the lighting, the face, the whole vibe is perfect, and then you look down and one hand has six fingers tangled like a knot of spaghetti. The instinct is to throw it away and reroll. Do not. There is a faster, smarter move that keeps everything you love and surgically repairs the one thing that is broken. It is called inpainting, and once it clicks you will never trash a good image over a small flaw again.
Hey friends. Let me describe the single most common heartbreak in this whole hobby. You run a batch, you scroll through the results, and one image stops you cold. It is the one. The composition sings, the light is gorgeous, the expression is exactly right. Then your eye drifts to a hand that looks like it lost a fight with a crab, or a background face that melted into a smear, or an earring that is somehow growing out of the cheek. So you sigh, hit generate again, and chase that magic for another forty rerolls without ever catching it twice.
Here is the move that ends that cycle. You do not have to regenerate the whole image to fix a small part of it. You can paint a mask over just the broken area and ask the model to redraw only that patch, blending it back into the picture you already love. That is inpainting. And its sibling, outpainting, lets you grow the canvas outward to give a cramped image more room to breathe. These two skills turn you from someone who prays for a perfect roll into someone who finishes images on purpose. Let me walk you through both, start to finish.
Inpainting means selecting a region of an existing image and regenerating only that region, while the rest of the picture stays frozen. You tell the tool, in effect, "everything is great except this spot, so leave the rest alone and just redraw here." The model looks at the surrounding pixels for context, reads a short prompt you give it, and fills the masked area with something that fits the lighting, the color, and the style of everything around it.
This is the same idea as the generative fill feature in modern photo editors, and it is built right into the major open tools. In a node workflow like ComfyUI, dedicated fill models such as FLUX.1 Fill handle inpainting and outpainting as a first-class task, and there are popular community workflows built specifically around repairing hands and faces. Whatever tool you use, the concept is identical: mask the flaw, prompt the patch, blend it in.
Think of inpainting as spot-repair, not a redo. You only touch the part that is wrong.
Here is the exact routine I run every time. It is short, and the order matters, because the two settings most people get wrong are mask feathering and denoise strength.
The single setting that fixes most bad inpaints: denoise strength, sometimes called inpaint strength. If your repair barely changes the broken area, your denoise is too low, so raise it until the model has room to truly redraw. If your repair ignores the surrounding image and pastes in something that clashes, your denoise is too high or your mask is too tight, so pull it back and give it more buffer pixels. Almost every "inpainting does not work for me" problem is really this one slider in the wrong place. Treat it as the first thing you adjust, not the last.
Hands and faces break for the same reason: they have tiny, high-stakes detail that the eye is ruthlessly good at judging. We are wired to spot a wrong finger count or an off eye instantly. The fix is the same inpainting routine, with a couple of tweaks per target.
Mask the whole hand, not just the bad finger, because a single weird finger usually means the underlying structure is off. Give the model the whole hand to rebuild. Keep the prompt anatomical and calm: "natural relaxed hand, correct fingers." Run it a few times, because hands are genuinely hard and your best result might be the third one. If the hand is small in the frame, an upscaling pass first can give the model more pixels to work with, which is why understanding your AI art upscaling pipeline pairs so well with inpainting.
For a melted background face or a slightly-off eye, mask the whole face region and prompt simply for "a clear, natural face, symmetrical eyes." Faces benefit from a slightly lower denoise than hands, because you usually want to preserve identity and expression while cleaning up the rough edges, not invent a brand-new person. This is also where your understanding of keeping a consistent character face across images pays off, since the same instincts that hold a face steady help you repair one without losing who it is.
| The flaw | How to inpaint it |
|---|---|
| Broken hand, extra fingers | Mask the whole hand + margin, higher denoise, prompt "natural relaxed hand, correct fingers" |
| Melted or smeared face | Mask the whole face, medium denoise, prompt "clear natural face, symmetrical eyes" |
| Weird object or artifact | Mask it, high denoise, prompt the clean surroundings so it gets painted over |
| Wrong small detail (jewelry, button) | Tight mask, medium denoise, prompt the exact thing you want there |
Now flip the idea around. Sometimes the image is not broken, it is just cropped wrong. The subject is crammed against the top edge, or you want a wide banner but generated a square, or the composition would sing with a little more empty space on one side. Outpainting solves this by extending the canvas beyond its original borders and letting the model invent what plausibly continues out there.
The workflow mirrors inpainting. You expand the canvas in the direction you want more room, the new empty area becomes the masked region, and the model paints a continuation that matches the existing scene. This is how you turn a tight portrait into a cinematic wide shot, or fix a composition where you accidentally cut off someone's feet. It is also a quietly powerful composition tool, because the extra breathing room you add can completely change how an image reads. If you have not thought much about where your subject sits in the frame, our guide to composition and framing will make your outpainting decisions a lot sharper.
Outpainting gives a cramped image room to breathe. Extend the canvas, let the model fill the edges.
Here is the mindset shift I want to leave you with. Beginners treat every generation as a lottery ticket: either the whole image comes out perfect or it goes in the trash. People who actually finish images treat generation as step one of a few. They expect to get something ninety percent of the way there, and they expect to spend a couple of minutes with inpainting and outpainting to carry it the last ten percent. That last ten percent is exactly where the broken hands, the smeared faces, and the cramped crops live, and it is the difference between an image that is almost good and one that is genuinely done.
You already know how to make a beautiful image. The thing that has been quietly costing you your best work is the belief that a single flaw means you have to start over. You do not. Mask the problem, write a tiny prompt, dial the denoise, and blend it back in. Extend the frame when the composition needs air. These two moves, inpainting and outpainting, are the patient craft underneath every clean, finished AI image you have ever admired, and they are completely learnable in an afternoon.
So next time you get the one, the image that almost made it, do not reroll. Zoom in on the part that is wrong, paint a mask over it, and fix it. Ninety seconds later you will have saved an image you were about to throw away, and you will start seeing your flawed-but-promising generations not as failures, but as drafts that are one quick repair away from being exactly what you pictured.