Why do AI hands get the finger count wrong?

A hand packs many joints into a small area and most training footage shows hands gripping, blurred or hidden. A hand with no named job guesses its own pose every frame.

Can I make the person look identical from clip to clip?

We have not measured this for any model and we do not promise it. A character reference and a short clip help. Check every result yourself.

Does more detail in the prompt help?

The right detail does: contact points, what the object is, a start and an end pose. A longer paragraph of mood words does not.

Which model is best for hands?

It depends on the shot. For hands that move, test Seedance 2.5 or Kling v3 against the model you started on; the pair above shows Seedance 2.5 on ours.

Will a restore pass fix a bad hand?

No. It sharpens what is there. Fix the structure first, then restore.

Guides / Video editing

Hands, faces and warping: the fixes that work

Hands, faces and rigid objects are where a clip gives itself away. There is no single switch that fixes them. There are six things that each help, and we tested every one on the same shot: a man adjusting a plain steel watch.

9 min read
September 24, 2026

Example

Reference

Restored

What you make, in order. Tap one to jump to its step.

Hands, faces and rigid objects break because they pack a lot of structure into a small part of the frame, and every frame is the model's fresh guess at it. Six things reduce the guessing: a prompt that names every contact point, a reference image, the model you pick, a shorter clip, a higher resolution and a restore pass at the end. None is a switch. Each removes one reason to guess.

We tested all six on one shot: a man in a black suit adjusting a plain steel watch, rendered on Seedance 1.5 Pro through the video composer. First, so you know what the problem looks like, a render of ours that got it wrong.

What goes wrong, and why

What you see	Why it happens	First fix to try
Extra or fused fingers, a hand holding nothing	The hand was never told what it touches	Name the contact: which hand, where, how many fingers
A bottle, a watch, a phone that changes shape	Rigid objects have no rigidity in the prompt	Say what the object is and give its proportions; add a reference image
A face that drifts during a turn	The face is redrawn from a fresh guess as it moves	Shorter clip, less head movement, a reference image; no promise about who it looks like
Soft or smeared fingers	Not enough pixels for the structure	720p or higher, then a restore pass

The failure specimen: a woman in an emerald blouse lifts a faceted glass perfume bottle in front of her face; the bottle changes its proportions while she holds it — **The failure specimen, not the example.** A render of ours from the same day: every finger count we checked is correct, and the bottle changes shape while the hands hold it. Watch the glass, not the face.

The bottle at 1.25 seconds, wide and squat — The same bottle one second apart. Drag the line across the glass. The prompt named each hand's grip and never said what the bottle was.

The bottle at 2.25 seconds, tall and narrow — The same bottle one second apart. Drag the line across the glass. The prompt named each hand's grip and never said what the bottle was.

The example: a watch that holds

This is the shot every fix below is tested on. Two hands, a rigid object, a push-in to the detail: the hardest kind of close-up. Five separate fingers on each hand at every moment we checked, knuckles and veins as the dial fills the frame, no lettering on the dial.

A silver-haired man in a black suit raises his left forearm and adjusts a plain steel watch with his right hand as the camera pushes in to the dial — **Made with Seedance 1.5 Pro** at 720p, 5 seconds, one run.

The prompt, exactly as sent

Premium watch advert, one continuous take, real-time. SUBJECT: a man of 35, silver-grey short hair, light skin, a black wool suit jacket buttoned over a black shirt, a plain steel watch with a black dial on his left wrist. HANDS: he raises his left forearm into the light and with his right hand adjusts the watch on his wrist, thumb and index finger on the strap, five separate fingers on each hand, knuckles visible. ACTION: the forearm rises into frame over the first second, the right hand settles the strap, then he turns the wrist so the dial catches the light and holds it. CAMERA: a slow push-in on a smooth dolly from a medium close-up to a close-up on the hands and the dial, eye level, no cuts. LIGHTING: one hard amber key from camera-left, a cold blue rim from behind camera-right on the jacket shoulder, a black background, haze. LOOK: glossy, anamorphic, fine grain, deep blacks, metal and fabric texture visible. No text, no logos, no brand name on the dial.

Tip:Why the hands held. The HANDS line gives each hand a grip and a finger count, and the object is named as a plain steel watch with a black dial. The one thing the prompt left open was the strap, and the strap is the one thing that changed between runs below.

Fix 1: name the contact and the object

Hands need a job and objects need a description. Take the HANDS line away and see what the model does with "he adjusts his watch".

The vague line

He adjusts his watch.

The watch shot with the hands line reduced to he adjusts his watch: the fingers hold, but a small mark appears on the dial — **Made with Seedance 1.5 Pro** at 480p. The fingers held in both. With the vague line the dial grew a small mark the prompt forbade, and the watch came back on a leather strap; with the contact named, a plain dial.

The same shot with each hand's contact named — **Made with Seedance 1.5 Pro** at 480p. The fingers held in both. With the vague line the dial grew a small mark the prompt forbade, and the watch came back on a leather strap; with the contact named, a plain dial.

Rigid objects get the same treatment. We added one line that says what the watch is and that it does not change.

The line we added

The watch is a rigid steel case, a round dial the width of his wrist, its diameter, its black dial and its strap width do not change at any point in the take.

The watch shot at 480p without the rigidity line — **Made with Seedance 1.5 Pro** at 480p. The case and dial held in both runs, so on this shot the line changed little. It costs nothing to keep, and the bottle above is what happens without it.

The same shot with the rigidity line added — **Made with Seedance 1.5 Pro** at 480p. The case and dial held in both runs, so on this shot the line changed little. It costs nothing to keep, and the bottle above is what happens without it.

Fix 2: give it a picture of the thing

A reference image replaces a guess with a target. Make a still of the object and add it under Reference images in the composer on a model that takes them. Seedance 1.5 Pro takes up to nine; address it in the prompt as @Image1.

Try it with our still

The watch reference we attachedWebP image · 1024x1024 · 73 KBDownload

Made with Seedream 4.5 from a one line prompt, cropped to the watch.

The line that binds it

@Image1 is the watch: the same plain steel case, the same black dial, the same proportions in every frame.

The watch shot without a reference: a steel bracelet — **Made with Seedance 1.5 Pro** at 480p with one reference image. The watch now matches the still: the black leather strap, the plain case, the dial. Two costs: the clip came back square although we asked for 16:9, and its first frames show the still itself.

The same shot with the watch still attached as a reference: a black leather strap, a plain steel case and the dial from the still, in a square frame — **Made with Seedance 1.5 Pro** at 480p with one reference image. The watch now matches the still: the black leather strap, the plain case, the dial. Two costs: the clip came back square although we asked for 16:9, and its first frames show the still itself.

Watch for this:A reference fixes appearance, not motion. The still says what the watch looks like, nothing about how a hand moves. Keep the contact line in the prompt too, and check the aspect ratio of what comes back.

Fix 3: try another model

Models fail differently. The same prompt on a second model is the cheapest way to find out whether the problem is the words or the model. For a hand that moves, the second model to try is Seedance 2.5 or Kling v3. Explore previews every video model by its own output.

The watch prompt on Seedance 1.5 Pro — **Same prompt, same seed, 480p, 5 seconds.** Seedance 2.5 kept five separate fingers on each hand through the adjust, pushed in to the dial, and turned the plain dial into the light at the end, on a steel bracelet with no lettering.

The watch prompt on Seedance 2.5: the man raises his forearm and adjusts a steel watch as the camera pushes in to a plain black dial — **Same prompt, same seed, 480p, 5 seconds.** Seedance 2.5 kept five separate fingers on each hand through the adjust, pushed in to the dial, and turned the plain dial into the light at the end, on a steel bracelet with no lettering.

Fix 4 and 5: shorter, and bigger

Drift has more time to build in a long clip, so render five seconds, then open the clip in your Library and press Extend clip to carry on from a frame you have already checked. On this shot the ten second run did not warp more; it spent the extra five seconds repeating the same adjusting gesture.

The watch prompt at 10 seconds: the adjusting gesture repeats before the push-in — **Made with Seedance 1.5 Pro** at 480p. Ten seconds held as well as five and cost twice as much; the model padded with a second adjust.

The watch prompt at 5 seconds — **Made with Seedance 1.5 Pro** at 480p. Ten seconds held as well as five and cost twice as much; the model padded with a second adjust.

Resolution is the other lever: fingers need pixels. These are two runs, not one run at two sizes, because on this model the seed does not reproduce a clip. What 720p buys is knuckles, veins and a dial you can read.

The watch prompt at 480p — **Made with Seedance 1.5 Pro.** The 720p run is the example at the top of the page.

The watch prompt at 720p — **Made with Seedance 1.5 Pro.** The 720p run is the example at the top of the page.

Fix 6: a restore pass, for what it is and is not

Restore Video renders a clip again at a size you pick, cleaning noise, compression and softness. It sharpens a finger; it does not add or remove one. Use it last, on the take that already has the right structure. Restore soft, noisy or old footage picks the engine by what is wrong with the clip.

A frame of the 480p clip scaled up to 720p — **Made with Restore Video** on the 480p clip, 4.4 seconds in. The same frame, scaled up on the left and restored on the right. The structure is the same; the edges are not.

The same frame after the restore pass at 720p — **Made with Restore Video** on the 480p clip, 4.4 seconds in. The same frame, scaled up on the left and restored on the right. The structure is the same; the edges are not.

Three ways to run this

Fastest: open the video composer, add the contact line, attach a still of the object under Reference images, and press Generate.
Hands-off: send the NOLGIA Agent the clip that went wrong and say what is wrong. Ask it to make a reference still of the object and render again with the contact named.
Most control: through the API send the same prompt to two models, attach the reference by asset id, and compare the job records.

Job	Use	Why
A hand or object that guesses	Prompt: contact and object lines	Removes the reason to guess, free
An object that must look one way	A reference image, up to nine on Seedance	A target instead of a description
Soft detail on a good take	Restore Video	More pixels on the structure you already have
One prop wrong in an otherwise good clip	Miroge, Swap	Keeps the shot and the move, changes the one thing

What it costs

Each fix is one more render, priced on the Generate button before you spend. Test at 480p and 5 seconds; render the keeper at 720p once. A restore pass is priced by the clip's length and the size you pick.

Seedance 1.5 ProPro and up
Per 5s clip (audio on by default)
480p 7 credits · 720p 15 credits · 1080p 33 credits
Seedance 2.5Pro and up
Per 5s clip
480p 26 credits · 720p 56 credits · 1080p 137 credits
1080p renders at 16:9 and 9:16

Read live from the catalog when this page loads. Every price is shown before you generate. See every rate.

Where you still regenerate

Two hands on one small object at once. Frame closer, or give each hand its own beat.
A face turning fast. Slow the turn in the prompt or cut before it.
A person who must look like a specific person. We have not measured this and do not promise it. A character reference helps; check the result yourself.
Glass and reflections. Anything seen through a rigid object belongs to the world behind it and is redrawn with it.

Checklist

Each hand has a job, including the idle one.
Each rigid object is named, with its proportions and, if it matters, a picture.
Five seconds, then extend. 720p for fingers in close-up.
Check the hardest detail first: if the hands hold, the rest usually does.
Restore last, never first.

Questions and answers

Why do AI hands get the finger count wrong?: A hand packs many joints into a small area and most training footage shows hands gripping, blurred or hidden. A hand with no named job guesses its own pose every frame.
Can I make the person look identical from clip to clip?: We have not measured this for any model and we do not promise it. A character reference and a short clip help. Check every result yourself.
Does more detail in the prompt help?: The right detail does: contact points, what the object is, a start and an end pose. A longer paragraph of mood words does not.
Which model is best for hands?: It depends on the shot. For hands that move, test Seedance 2.5 or Kling v3 against the model you started on; the pair above shows Seedance 2.5 on ours.
Will a restore pass fix a bad hand?: No. It sharpens what is there. Fix the structure first, then restore.

Give the hands a job in your next clip

Open the video composer

Prompting

Seven video prompt mistakes, and the fix for each

The seven ways a video prompt goes wrong, each rendered twice on Kling v3: the vague prompt, then the same prompt with one fix. Copy the fixes, and read what a seed does not do.

September 24, 2026 · 11 min read

Video editing

Restore soft, noisy or old footage

Restore Video re-renders a clip from your Library at up to 8K, cleaning noise, compression and softness. Pick the engine by what is wrong with the clip, and know what it cannot bring back.

September 24, 2026 · 6 min read