Extra arms and duplicated limbs in wide shots

Full-body wides and group shots are where extra arms cluster. Not because the model does not know how many arms a person has, but because each body occupies too little of the frame for the pose to be decided. The model averages across an enormous pose space and lands on something approximately a person: a spare forearm, a second pair of hands, a leg that belongs to the extra on the left. Re-rolling the whole wide trades a frame that is mostly right for a fresh draw on every body in it. The composition you liked will not come back.

Treat the extra limb like a bad hand, at a larger scale. Mask it, give the region enough pixels to be a body, and only then decide whether you also need to supply the pose. Fixing hands without regenerating the image is the same loop aimed at fingers. This is the version for arms, legs, and the fused torso that shows up when two people overlap.

Extra arms are a pixel-share problem

A hand in a medium shot might be two percent of the frame. A whole body in a full-length wide, sharing the canvas with a second body and a room, can be not much more. Training objectives average error across the image, so a region that small can be confidently wrong while the picture still scores well. The hero's face can be excellent and the left arm a rumour.

Two things make limbs worse than a wall in the same pixel budget.

Articulated occlusion. Limbs hide each other. In a group, they hide other people's limbs too. Averaging across that produces fused elbows, a shared wrist, an arm that reads as belonging to both people at once.

Pose variety. There is no single "standing" the way there is a relatively stable "front face." Weight on which foot, which elbow is bent, whether a hand is in a pocket: each is a different picture. Under-specified, the model blends two legal poses into one illegal body.

So duplicated limbs show up most in full-body shots where the figure is small, in two-shots where bodies overlap, and in wides that also ask for tiny hands and a readable face. If the extra arm is a copy of the other arm, or a second torso appears at the edge, the canvas is tiling. Drop the size. The loop below is for the local case: one extra forearm on an otherwise single, well-scaled figure.

Don't re-roll the frame

The reflex is to run the prompt again. Do not, if the rest of the picture is the one you would have art-directed for an hour. You are re-sampling every extra, every hand, every bit of set dressing, in the hope that the limb count comes out legal. Re-roll when several bodies are illegal at once, when the extra limb is fused with a silhouette you cannot mask apart, or when the canvas itself is tiling. Otherwise, edit.

Inpainting is the tool: you draw a boundary, the model repaints inside it, everything outside is preserved exactly. Element re-render versus masked inpainting is the decision between naming "the extra arm" and selecting it. In a two-shot, "the arm" is ambiguous. Mask.

Masked inpaint that actually lands

Crop in so the broken body is most of the frame. An extra arm that is two percent of a wide will still be two percent of the model's budget if you inpaint at the full canvas. Crop to the person plus a margin of the set, edit at that scale, composite back.
Mask the illegal anatomy plus everything it is fused to. A mask that covers only the ghost forearm and stops at the invented wrist produces a correct-looking arm attached at a point that was never a shoulder. Include the shoulder, a slice of torso, the sleeve, and the other arm if the two are interacting. Feather the edge.
Prompt the pose, not the noun. "A person with two arms" names the concept and does not constrain the joints. Describe the geometry: which arm is visible, where it is, what it is doing.
Iterate on the mask and the pose prompt. Hosted image-to-image has no denoise dial. Ghost limb: widen the mask to the joint. Whole body changed: tighten the mask and keep the torso.
On the agent call, pin seed while you iterate on wording. Changing both at once is how you spend six attempts learning nothing.
Stop at three. If the third attempt is still illegal, the pose is under-determined. Supply geometry (next section) or change the shot.

Weak	Strong
"fix the extra arm"	"one left arm, elbow bent, hand in the front jeans pocket, sleeve to the wrist, torso otherwise unchanged"
"normal human body, two arms"	"standing weight on the right foot, left arm hanging at the side, right hand on the strap of a bag on the right shoulder, five fingers, no other limbs"
"remove the third arm"	"the space under the left elbow is the dark jacket and the wall, no limb in that gap"

"Remove the third arm" still names an arm. Describe the pixels that should occupy that region: jacket, wall, empty air. Run the edit from the photo editor: an instruction plus a mask when the sentence is ambiguous. On the agent path, generate_image_from_image takes mask and seed; mask is honoured only by mask-aware edit models.

On video, do not inpaint thirty frames by hand. A spare arm that is only a little wrong in frame one will reshape itself across the clip. Repair the still and generate image-to-video from the repaired plate. Keep clips short enough that a bad limb does not have six seconds to evolve.

Pose-conditioning when the body is the problem

Three failed inpaints means the model is still guessing the pose. Stop describing it and start supplying it.

A still of a real pose, as a reference. Photograph yourself or a colleague in the stance you want, similar camera height, similar light. Feed that as a reference image on the masked edit. You have collapsed the pose space to one that already has two arms in the right places. Keep the reference boring: one person, readable silhouette. A reference lit from the opposite side will fuse.

Motion control, when the shot is a clip and the pose is a performance. Motion control takes a driving video and an image of the subject, and re-renders the performance on that subject. A clean full-body driving clip with legal limb count is pose-conditioning across time. Kling V3 Pro motion control is the catalog's pro-tier entry. It will not fix a driving clip that already has a spare arm. It will copy a legal pose onto a legal still.

Open-model pose control, when you own the stack. The established stills path is a pose map (OpenPose-style skeletal conditioning) fed in as structure so the model paints inside a skeleton it did not have to invent. That is a local ComfyUI-class workflow, not a named toggle on the hosted form. Treat it as the escalation when you already work that way.

If none of that is worth the time, change the shot. A full-body wide of two people overlapping with empty hands is close to the worst case. Give each body something to hold, separate them in depth, or recrop to a medium. Hands, teeth and signage already recommend giving a hand a rigid object; the same constraint helps the arm it is attached to.

Shot design that starves the failure

Stop at the waist unless the feet are the point. Medium shots give the torso and the face a budget. Full-body is a request for hands, feet, garment hem and pose, all at once.
Separate overlapping bodies. Shoulder-to-shoulder two-shots fuse limbs. Put one person a half-step forward, or cut to singles.
Put something in the hands. A cup, a phone, a bag strap. Empty, mid-gesture arms are where extras grow.
Do not also ask for a crowd of readable faces. A wide that must solve extras' anatomy and extras' faces is two pixel-share failures in one frame.
Check limbs before you check taste. Look at armpits, waist, and the gap between two people.

Repair a still before you animate it. Image-to-video will commit to the silhouette it is given. A ghost arm in frame zero is a ghost arm with motion.

FAQ

The inpainted arm looks pasted on. What did I miss?

The mask boundary. A hard edge leaves a seam; a mask that also ate half the torso will relight the join. Feather, include the sleeve so the join happens inside fabric, and check that the new arm is lit from the same side as the shoulder you kept.

Should I mask the extra limb or the whole person?

Mask the illegal region plus the joints it attaches to, not the whole figure, if the legs and the face are already right. Whole-person inpaint is a re-roll wearing a mask. If the silhouette itself is wrong — fused with another person, or a duplicate torso — recrop the shot.

Why does a group shot keep growing arms even at high quality?

Quality rungs add pixels to the file, not budget to each body. Six people in a wide are still six small bodies. Split the group, recrop to the two you actually need, or push extras out of focus. A legal two-shot over a bokeh crowd will beat a "sharp" group with spare elbows.

Can I prompt "anatomy correct" in the negative and skip the mask?

You can put it in the prompt. You should not skip the mask once the extra limb exists. Negation is bad at erasing a structure that is already in the pixels. On a fresh generation, write the pose. On a broken frame, mask.