Fixing a bad hand in video, not a still

Exporting the worst frame, inpainting the hand, and dropping the still back onto the clip is not a video fix. It is a still fix glued to 25 moving pictures a second. The corrected fingers exist for one frame. Every neighbour still has the fused knuckle, and now you also have a one-frame pop where the geometry disagrees.

A still has to be right once. A clip has to be right on every frame, and the failure that reads as "AI" is the pose changing. A hand that is merely soft in frame one can reshape itself by frame thirty. That is temporal consistency on top of the anatomy problem, and it is why the still-image recipe from fixing hands without regenerating the image does not transfer.

Three options work on a moving hold: a masked video edit, a short re-generation of just the bad interval, or a reframe that takes the hand out of shot. Pick on how long the defect lasts and whether the hold is the point of the shot.

Why per-frame inpaint fails

Versely renders at 25 frames per second by default, so a five-second clip is 125 frames. Each frame is a fresh sample. Nothing in a diffusion video model enforces pixel identity for a hand across that span.

A single corrected frame is a flicker, not a repair. Inpaint every bad frame independently and you are drawing a new hand 125 times; the count can be five on each still and still boil in motion. The contact line is also a moving occlusion. Fingers wrap and hide the product as the hold tightens, so a still mask does not track.

The job is to constrain the same region across time, or stop showing that region. Judging the result on a paused frame will lie to you. Play it.

Option 1: Mask the hand across the clip, then edit that region

Use this when the rest of the shot is a keeper (lighting, face, set, timing) and the defect is localised to the hand, or the hand-plus-product contact, for most of the duration.

The sequence is tracked, not per-frame:

1. Isolate the offender. Sam 3 Video Segment is selection, not generation. The prompt names the object ("the right hand," "the hand on the mug"). When words are ambiguous, add a box or click points so it does not grab the other hand or the sleeve. You get a mask that follows the region through the clip, which is how you confirm you are about to edit the right thing.
2. Edit with an instruction, not a scene description. Wan 2.7 Video Edit takes the source clip, a prompt that states the change, an optional reference image, and an audio_setting. Write the hold, not the shot. Name the hand so the instruction cannot wander. Set audio to origin if the original soundtrack has to survive; leaving it on auto hands the model the decision and you can lose a voiceover you were not trying to replace.
3. Give it a real hold as a reference if you have one. A still of a correctly counted hand, at a similar angle, in similar light, or a packshot of the product being held. The reference is there so the model is not inventing the pose from text on every frame.

A usable instruction:

Replace the right hand with a natural human hand, five fingers, thumb on the mug handle, fingers wrapped around the ceramic. Keep the mug, the sleeve, the face and the background unchanged. No extra fingers, no fused knuckles.

What that instruction is not: a new scene. The source video already is the scene. Re-describing it is how you get a different shot with a different hand.

Wan 2.7 Video Edit takes 0 for the full source, or 2 to 10 seconds for a trimmed piece. Prefer the trimmed piece. A three-second repair has less time to drift than an eight-second one.

Play the result at full speed and look at the contact, not the fingers in isolation. A counted hand that slides around the mug is still a fail. If the product deforms as the fingers attach, keep the visible SKU outside the mask.

Option 2: Re-generate only the bad interval

Use this when the defect is concentrated. The hold is fine for five seconds, then the hand reshapes as it sets the bottle down. Re-editing the whole clip to save one second is the expensive move.

Cut the interval out, generate a short replacement, and join it.

1. Cut on motion, not on the ugly frame. Take the section out with a little handle on each side so the replacement has overlap to hide the join. A cut in the middle of a grip, with no shared pose, is a teleport.
2. Keep the replacement short. Do not extend a bad clip; replace the segment. Veo first-and-last-frame generates at 4, 6, or 8 seconds — use 4s for a repair. Drift has less room to run in the shortest clip the model will actually make.
3. Anchor both ends. Feed a clean frame from just before the cut as the first frame of the replacement, and a clean frame from just after as the last, so motion has to travel through the join rather than jump.
4. Match the camera. One move, stated at the end of the prompt, or "static, locked-off, tripod" if the original was locked. A replacement that introduces a push-in on a locked shot will not join.

Judge the join in the editor. The 480p preview pass is free and carries a short per-user cooldown; a single charge applies to the final export only, regardless of clip count. Check the cut at preview resolution before you pay for a finish you have not seen in motion.

If the replacement still reshapes the hand, the pose is a bad ask. That is option 3.

Option 3: Reframe the hand out of the shot

Use this when the hold is not the brief, or when two repair attempts have produced a different wrong hand. Some poses are not worth fixing. A mid-gesture empty hand, fingers spread, seen edge-on, is close to the worst case in stills and worse in time, because the occlusion pattern changes every frame.

Three reframes, in order of how much they cost:

Crop. A tighter frame that loses the hand and keeps the face, the product on the table, or the label. An aspect-ratio change is a geometry decision, not a generation.

Cut away before the hold completes. The riskiest moment is often the static pose at the end of a reach, not the reach itself.

Prompt the next take without a hero hand. Product on the surface, hand in the blurry foreground, or the hold off the focal plane. Motion and shallow depth of field over a mid-gesture hand read as camera work.

This is the option people resist because it feels like giving up on a clip they have already paid for. It is usually the cheapest of the three. Repairing a hold you did not need is how a small defect becomes a second full generation plus an edit pass.

Which of the three to run

What you see	First move	Stop and switch if
Rest of the shot is a keeper; the hand is wrong for most of the duration	Masked video edit (option 1)	The product warps, or the hand is still boiling after one edit
The hand is fine, then fails for a second or two	Segment retake (option 2)	The join teleports, or the replacement introduces a new camera move
The hold is not the brief, or two repairs have failed	Reframe (option 3)	(This is the exit. Do not repair a third time.)
Empty hand, mid-gesture, edge-on, small in frame	Reframe first	You already know the pose is a bad ask

Play the candidate at full speed before you keep it. A paused frame that looks counted will still fail if the knuckles crawl. Do not "help" a repair with a heavy upscale: over-sharpening amplifies the flicker the edit was meant to hide.

If you are still generating rather than repairing, keep the clip short, give the hand a rigid object, and treat a hero grip as a shot you may have to take apart.

FAQ

Can I inpaint one frame and let it stand in for the clip?

No. A still inpaint is correct for one picture. On a clip it is a one-frame geometry change, which reads as a pop, and it does nothing to the neighbouring frames at 25 fps. If the defect is a single-frame flash, cut it. That is an edit, not an inpaint.

How short should a segment retake be?

Four seconds when you are using first-and-last-frame, which is the shortest that model offers. A one-second insert is hard to join; an eight-second replacement is just another full generation. Anchor both ends with real frames from the source.

Does the video mask need to include the product?

Only the contact the hand covers, and only if you will let the model rebuild that contact. If the SKU must stay exact, keep the visible product outside the mask. A counted hand on a melted label is not a fix.

Why not just re-roll the whole clip?

Because you would be re-rolling the parts that already work. Localise the defect (mask or segment), or hide it (reframe). Whole-clip regeneration is for when the shot itself is wrong.