Fixing hands without regenerating the image
Fused fingers are a small-region problem, not a whole-image one. A masked inpaint loop that lands in two or three tries, plus the depth-guided fallback.
You have an image where everything worked. The light is right, the expression is right, the composition is one you would have had to art-direct for an hour to get on a shoot. And the left hand has five fingers plus something that might be a sixth, fused at the second knuckle.
The reflex is to run the prompt again. Do not. You are trading a frame that is ninety-eight percent correct for a fresh draw on the whole distribution, and the odds that the next one gets both the composition and the hand are worse than the odds of fixing the hand where it stands. The hand is a small-region problem, and small-region problems have small-region solutions.
Why hands fail, mechanically
Two things are working against you at once.
Pixel share. A hand in a medium shot might be two percent of the frame. Training objectives average error across the image, so a region that small contributes very little to the loss, and a model can be confidently wrong there while scoring well overall. There is no mechanism telling it that humans scrutinise hands more than they scrutinise a wall.
Articulated occlusion. A hand has many joints, an enormous space of plausible poses, and fingers that constantly hide each other from the camera. Averaging across that pose space produces something approximately hand-shaped rather than a specific, physically consistent one. That is where the extra finger comes from: not ignorance of what a hand looks like, but a blend of several correct hands into one that is not.
Newer image models have narrowed this considerably. They have not closed it, and the failures that remain are the ones where both problems compound — a small hand in an unusual pose, partially behind something.
The useful consequence is that the fix follows directly from the cause. If the problem is that the region got too little of the model's budget, give it more of the budget.
The loop that usually lands in two or three tries
- Crop in before you edit. This is the step people skip and it is the one that does the work. Crop to the hand plus surrounding context, edit at that scale, then composite back. You have taken a region that was two percent of the frame and made it most of it, which changes the pixel-share problem from the cause of the failure into a non-issue. Many editing surfaces do a version of this automatically; doing it explicitly is more reliable.
- Mask the hand plus a margin of everything it touches. A mask that stops exactly at the wrist produces a correct hand joined badly to the arm. Include the cuff, the last centimetre of forearm, and whatever the hand is holding. Feather the boundary.
- Write the pose, not the noun. "A hand" gives the model the same ambiguous pose space that failed the first time. Describe the geometry: which hand, how many fingers are visible, where each contact point is.
- Set denoising strength around the middle and adjust from what comes back. Too low and the fused fingers survive under a slightly different texture. Too high and you get a new hand that no longer matches the arm's angle or the sleeve. This single dial explains most edits that come back either useless or unrecognisable.
- Fix the seed while you iterate on wording. Changing the phrasing and the seed together means you cannot tell which one moved the result, and you will iterate blind for six attempts.
- Stop at three. If the third attempt is still wrong, more attempts are not going to converge — the pose is under-determined and you need to supply geometry rather than hope. Go to the next section.
On the prompt itself, the difference between phrasings is larger than most people expect:
| Weak | Strong |
|---|---|
| "fix the hand" | "a relaxed left hand, four fingers curled around the mug handle, thumb resting on the rim, back of the hand toward camera" |
| "correct fingers" | "five fingers, index and middle extended, ring and little finger folded into the palm, seen from the side" |
| "natural hand" | "an open right hand at rest on the table, fingers slightly apart, palm down, knuckles catching the window light" |
Note what the strong versions all do: they state a count, name the contact points, and give the camera's view of the hand. Count and contact are what constrain the pose. View is what stops the model solving a correct hand at the wrong angle for the arm it is attached to.
The masked route is available directly — Versely's image-to-image tool takes a mask alongside the prompt and the seed, so this is the same loop whether you drive it from the photo editor or by asking the agent to edit a photo. Element re-render versus masked inpainting is worth reading first if you are choosing between masking the hand and naming it, because hands are a case where the mask usually wins: "the hand" is ambiguous the moment there are two of them in frame.
When three tries have not landed
At that point you stop describing the pose and start supplying it.
Give it a reference. Photograph your own hand on a phone at the same angle, in roughly the same light, and pass it as a reference image alongside the masked edit. This is the single most practical version of "supply the geometry", it takes about ninety seconds, and it works because the model no longer has to resolve a pose from words. Keep the reference clean: one hand, plain background, similar lighting direction to the target frame. Mismatched conditions across references produce fusion errors, where the model averages two inputs into a third thing.
Or go structural, in an open-model stack. The established fix in the Stable Diffusion and Flux lineage is depth conditioning with a hand-specific preprocessor — ControlNet depth driven by a HandRefiner-style step that estimates a plausible hand mesh and feeds its depth map in as structure. The model then paints inside a geometry it did not have to invent. This is a ComfyUI-class workflow rather than something you drive from a hosted generation interface, so treat it as the escalation path when you own the pipeline, not as the default.
Or change the shot. Sometimes the honest answer is that the pose is a bad ask. A hand mid-gesture, fingers spread, seen edge-on with nothing to hold is close to the worst case. Giving the hand a rigid object anchors the pose space, and that is a smaller change than four more failed inpaints.
Prevention is cheaper than the loop
- Give hands something to hold. A cup, a phone, a railing. Constraining the pose space is the whole point.
- Generate at a resolution where hands get real pixels. A hand at 512 pixels across is a different problem from a hand at 120. Two-pass regeneration versus native high resolution covers when generating large natively beats generating small and upscaling, and hands are one of the clearest cases for going native.
- Frame hands away from the fine-detail budget or out of frame entirely where the shot allows it. Hands, teeth and signage has the full shot-design version of this.
- Check hands first when reviewing a batch. They are the fastest disqualifier, so looking at them before you evaluate anything else saves you from falling in love with an image you will have to fix.
Video is a different job
A hand that is merely soft in frame one can visibly reshape itself by frame thirty, because the per-frame problem now compounds across the clip. Fixing a single frame does nothing, and inpainting thirty of them by hand is not a workflow.
The path that works is tracked rather than per-frame: isolate the hand across the clip with segmentation, then either composite the fix against that mask or run a prompt-driven edit over the clip. Sam 3 Video Segment produces the mask; Wan 2.7 Video Edit modifies a source video from a prompt plus an optional reference image. It is more setup than a still-image inpaint, which is a strong argument for the shot-design fixes on anything that moves — and for keeping clips short, since a three to five second clip has less room to drift than an eight second one.
FAQ
Why does my inpainted hand look pasted on?
Almost always the mask boundary or the strength. A hard mask edge leaves a visible seam, and a strength high enough to produce a good hand will also produce one lit slightly differently from the rest of the frame. Feather the mask, include the cuff and forearm so the transition happens inside a region with texture, and check the light direction on the new hand before accepting it.
Should I fix the hand or re-roll the image?
Fix it, unless the composition was mediocre anyway. A reroll changes everything, and everything includes the parts that were already right. The exception is when the hand failure is a symptom — if the pose is anatomically implausible for the arm it is attached to, the model was confused about the body, and no regional edit repairs that.
Do any models just get hands right now?
Current image models are meaningfully better than the generation that produced the six-finger meme, and simple poses in good light are usually fine on the first try. Difficult poses in small frame regions still fail. The ranked view of editing models is the practical way to compare what is available for the repair pass, with credit costs shown beside each.
Is "extra fingers, deformed hands" in the negative prompt worth keeping?
Not as a standing incantation. It occupies prompt attention on every generation to address a defect that appears in a minority of them, and on models that run without classifier-free guidance the negative field may do nothing at all. Repair the failures you actually get instead of taxing the ones you do not.