Your reference image has clean, anatomically correct hands. The first frame of the generated clip still looks fine. Then the character reaches for a cup, waves, or just walks past the camera, and the hands smear into the sleeve, fuse with the torso, sprout a sixth finger, or vanish entirely for 8 frames before snapping back. This is one of the most reliable failure modes in current AI video: hands are small, articulated, fast-moving, and self-occluding, which is exactly the region where a diffusion model has the least signal to work from.
Fastest fix: name the hand explicitly in the prompt (both hands visible, five fingers on each hand), reframe so hands fill at least 8 to 10 percent of the frame, and keep continuous hand action under ~2 seconds — split longer takes with a cut. If the clip is already rendered, mask just the hand region and regenerate it (Runway Inpainting or Kling Inpainting) instead of re-rolling the whole shot.
This is a model-class limitation, not a bug in your account. As of June 2026, hands during motion still break across every frontier model — Sora 2, Veo 3.1, Kling 3.0, Runway Gen-4.5, Hailuo, Pika, Seedance — though Veo 3.1 and Kling hold hand geometry noticeably better than the rest. The fixes below are about working around the limitation, not waiting for it to disappear.
This article covers why hands break specifically during motion (not stillness), how to phrase prompts that minimize it, how to confirm the fix worked, and how to recover a shot when re-rolling is not an option.
Which bucket are you in?
Pause the clip frame-by-frame and match what you see to the most likely cause. This points you at the right fix below instead of re-rolling blindly.
| What you observe | Most likely cause | Go to |
|---|---|---|
| Hands fine when still, smear only mid-motion | Hand too small / motion blur threshold | Steps 1, 2 |
| Six fingers / fused fingers appear | Prompt never names the hand | Step 1 |
| Hand vanishes only when it crosses body or same-tone background | Segmentation failure | Step 3 |
| Hands wrong from the moment they enter frame (image-to-video) | Hidden-hand starting frame | Step 4 |
| Fingers warp only while gripping an object | Grip geometry conflict | Steps 1, 6 |
| First ~2s correct, then progressive drift | Take too long | Step 5 |
| Hands stretched on a wide / fisheye / GoPro shot | Lens spec amplifies near-camera distortion | Cause 7 |
Common causes
Ordered by hit rate, highest first.
1. Hand is small relative to frame and crosses the motion blur threshold
When a hand occupies less than ~3% of the frame and moves faster than the model’s training distribution allows for that resolution, it gets aliased into a blur that the decoder cannot reconstruct as an anatomically valid hand. The model picks “looks like a sleeve” over “looks like a half-resolved hand.”
How to spot it: Pause every 4 frames. Hands are intact when stationary, degrade only during the motion segment, and recover when motion stops.
2. Prompt focuses on the action, not the hand
Prompts like “person waves hello” or “barista pours coffee” describe the action. The model interprets the action holistically and treats the hand as a means to the verb, not a region to preserve. Hands get optimized away in favor of the dominant motion vector.
How to spot it: Your prompt names the verb but never the noun “hand” or “fingers.” A prompt that explicitly mentions five visible fingers gives the model a region to defend.
3. Hand crosses in front of a similarly-colored region
Hand passes in front of the face, the torso, or a same-tone background. The model’s segmentation between hand and background fails for a few frames, and the hand visually fuses into whatever it crossed.
How to spot it: The disappearance happens exactly when the hand overlaps a same-color region. Move it across high-contrast space and the bug stops.
4. Image-to-video extension of a closed-fist or hidden-hand starting frame
If the starting image hides the hands (pockets, behind back, closed fist), the model has no anchor for finger count, knuckle position, or palm orientation. The moment hands enter frame, the model invents them from scratch, often badly.
How to spot it: Reference frame has hands hidden, and the artifact appears precisely when hands first become visible.
5. Holding an object — fingers wrap the object incorrectly
Pen, cup, phone, steering wheel — anything the hand grips. The model has to simultaneously render correct grip geometry and consistent object size. It usually fails at one. Fingers pass through the object, the cup floats, or the pen warps.
How to spot it: Hands look fine when empty; fail only when grasping. The object’s shape distorts in sync with the finger errors.
Counterintuitively, grip is often easier to stabilize than free-air gesture. A hand anchored to a rigid object has fewer valid poses for the model to choose from, so naming the contact explicitly (fingers firmly grip the edge of the ceramic cup) constrains the geometry. Vague free-hand motion (she moves her hands) gives the model the most room to drift — this is Kling’s own published advice as of June 2026.
6. Motion segment exceeds ~2 seconds of continuous hand action
Most current models stay anatomically stable for ~1.5–2 seconds of complex hand motion, then drift. Long takes with continuous hand work (typing, sign language, gestures) accumulate error.
How to spot it: Hands are correct for the first ~40 frames, then degrade progressively. Shortening the clip eliminates the issue.
7. Wide-angle lens spec amplifies hand distortion
“Wide-angle lens,” “fisheye,” or “GoPro” in the prompt teaches the model to exaggerate near-camera elements. Hands closest to the lens get stretched into the distortion budget, which models render as anatomical drift rather than honest perspective.
How to spot it: Removing the lens spec while keeping every other prompt term fixed produces normal hands.
Shortest path to fix
Step 1: Add explicit hand language to the prompt
Don’t just describe the action. Add structural anchors:
"a barista pours espresso, both hands visible,
five fingers on each hand, fingers wrap naturally
around the cup handle, hands occupy lower-third of frame"The phrase “five fingers” alone reduces extra-finger artifacts in most models because it gives the denoiser a count to honor. Diffusion models understand “fingers” as a concept but have no built-in counting mechanism, so telling them what correct looks like in the positive prompt works better than only listing what to avoid.
If your tool has a separate negative-prompt field (Kling, Hailuo, and most ComfyUI/local pipelines do; Sora and Veo do not), add a short one: extra fingers, fused fingers, deformed hands, mutated hands, missing fingers. Keep it to roughly five terms. Over-stuffing the negative prompt past ~5 hand-related terms is a documented failure mode in diffusion models: beyond that threshold outputs get sterile and, paradoxically, more artifact-prone (inverse amplification). Lead with your single biggest problem term, e.g. extra fingers first.
Step 2: Keep hands large enough in frame
Reframe the shot so hands occupy at least 8–10% of the frame area during the motion segment. Medium shot beats wide shot for any clip where hand motion is the subject.
If you can’t reframe, generate at higher resolution (1080p → 4K) and downscale. The model has more pixels to spend on hand detail at higher resolution.
Step 3: Avoid hand-overlap-with-similar-tone regions
If the action requires the hand to cross the body, change one of:
- Wardrobe color (high-contrast sleeve vs. background).
- Hand position (cross higher or lower to avoid the torso midline).
- Lighting (rim light separates hand from background).
Step 4: Anchor with a hands-visible starting frame for image-to-video
If you’re doing image-to-video, the reference frame must show the hands you want, in the position they’ll start from. Closed fists, pockets, or behind-back starting poses are the single biggest predictor of hand drift in the generated motion.
Step 5: Shorten and stitch
Split a 4-second hand-heavy clip into two 2-second clips with a cut. Each shorter generation will hold anatomy better, and a clean cut between them is invisible if the action continues. Avoid asking one model pass to hold hands for more than ~2 seconds of continuous action.
This matters even though most models advertise much longer clips. As of June 2026, the per-clip ceilings are: Veo 3.1 8s, Runway Gen-4.5 ~16s, Kling up to ~2 minutes, and Sora 2 15s on ChatGPT Plus or 25s on the web Storyboard for $200 Pro accounts. The advertised maximum is not the anatomically-safe maximum. Hand stability degrades long before the clip-length cap, so generate the hand-heavy beat as its own short shot regardless of what the model will let you request.
Step 6: Mask-and-regenerate just the hand region
If re-rolling the whole clip isn’t viable, mask the hand region only and regenerate that area while keeping the rest of the clip frozen. This costs less than a full re-roll and preserves the parts that worked.
- Runway — open the clip, choose Inpainting, brush over the hand to create a mask, then describe the replacement (
five-fingered hand, natural anatomy, holding cup). Runway regenerates the masked region frame by frame. - Kling — use Inpainting (Kling 3.0), paint a tight mask over just the hand, write the corrective prompt, and set Redraw Intensity low-to-mid so the model respects the surrounding pixels instead of reinventing the whole region. Tight masks are the rule for detail work like hands, jewelry, and faces.
- Local / ComfyUI pipelines — the MeshGraphormer Hand Refiner node is the current state-of-the-art for hands (June 2026): it estimates the hand’s depth and 3D mesh and rebuilds it, which removes the “melted candle” look that a plain inpaint can leave behind.
For a manual inpaint, these settings are a good starting point:
Mask: hand bounding box + 20px feather
Prompt: "five-fingered hand, natural anatomy, holding cup"
Strength / redraw intensity: ~0.7 (keep some motion from the original)Step 7: Hide the failure with intentional motion blur or cut
If all else fails: add post-production motion blur over the bad frames (radial blur centered on the hand), or cut to a different angle for the 8 to 12 frames where the hand breaks. Audiences forgive a cut; they do not forgive a six-fingered hand on screen for half a second.
How to confirm it’s fixed
Don’t trust the thumbnail or the first frame. Verify the way an editor would:
- Scrub frame-by-frame through the whole motion segment, not just play it once at speed. Most hand breaks last 4 to 12 frames and are invisible at normal playback.
- Pause specifically at the moment the hand crosses the body, enters frame, or grips the object — the highest-risk instants from the diagnosis table.
- Count fingers on every paused frame in the motion segment. A clip that holds five fingers at rest can still flash six mid-reach.
- Loop the segment once at 0.25x speed. If nothing pops at quarter speed, it will hold at full speed for viewers.
A shot passes when you can scrub the entire hand-motion segment and never see a finger count change, a fused finger, or a hand-into-sleeve smear.
When this is not on you
Hands during motion are a known weak point across every current frontier video model — Sora 2, Veo 3.1, Runway Gen-4.5, Kling 3.0, Hailuo, Pika, Seedance (June 2026). Even Google’s own benchmarks frame this as incremental: Veo 3.1 improved frame consistency roughly 40 to 60 percent over Veo 3.0 across 8-second clips, which is real progress and still not “solved.” Some shots — sign language, juggling, or hands-only close-ups during fast motion — are simply not yet achievable with one-shot generation. Plan around the limitation rather than re-rolling into it.
Easy to misdiagnose as
- “Bad seed.” Re-rolling rarely fixes hands-during-motion; it just shuffles which frames break. Address the cause, not the variance.
- “Model is bad.” Hands break across models in the same class of motion. Switching models without changing the prompt or framing usually reproduces the problem.
- “Prompt is wrong.” The prompt may be fine; the issue is often framing, duration, or starting-frame visibility — none of which the prompt alone controls.
Prevention
- Default to medium shots for any clip with prominent hand motion.
- Add “five fingers on each hand, both hands visible” to your hand-motion prompt template.
- Keep continuous hand-action segments under 2 seconds; stitch longer takes.
- If starting from a reference image, never use a hidden-hand starting pose.
- Build a “hand-safe” prompt module you reuse across all character clips, separate from the action description.
- Where the action allows, anchor hands to a rigid object instead of letting them gesture in open air — a constrained grip drifts less than free motion.
- For client work, plan a cut-away shot at the hand region as fallback B-roll.
FAQ
- Which model has the best hands during motion right now? As of June 2026, Veo 3.1 and Kling 3.0 hold hand geometry the most reliably in independent testing; Sora 2 is strong on identity and expression but still drifts on small physical details (hands, shoelaces) unless you anchor with a reference image. There is no model that “solves” hands during fast motion — switching models is a tweak, not a fix.
- Why does re-rolling the same prompt rarely fix it? Re-rolling only changes the seed, which shuffles which frames break, not whether hands break for that shot. The cause is framing, duration, or starting-frame visibility, none of which the seed controls. Change a cause from the table, not the seed.
- Why are hands worse than feet? Hands are more articulated, move faster, occupy more frame attention, and have more training-data variance (different sleeves, gloves, accessories). Feet are usually static, partially occluded by floor or pants, and forgive distortion.
- Does upscaling fix hands? Upscaling sharpens existing pixels but cannot invent correct anatomy. If the hand is broken at 720p, it will be broken at 4K. Fix the generation first, then upscale.
- Do negative prompts for hands actually help? Only where the tool exposes a negative-prompt field (Kling, Hailuo, ComfyUI — not Sora or Veo), and only in moderation. Keep it to about five hand terms; past that, diffusion models tend to produce more artifacts, not fewer. A clear positive prompt (
five fingers, anatomically correct hands) usually outperforms a long negative one. - Should I just fix the hand in post instead? Often yes. For a single bad beat, masking the hand region and inpainting it (Step 6) or cutting away for 8 to 12 frames (Step 7) is faster and cheaper than re-rolling a whole clip and praying the seed cooperates.
Related
- AI Video — Face Changes Mid-Clip
- AI Video — Character Drift Across Clips
- AI Video — Product Deformation During Motion
- AI Video — Unnatural Physics in Object Interaction
- AI Video — Motion Incoherent
- AI Video — Multi-Character Swapped Mid-Clip
- AI Video — Image-to-Video Not Following the Reference
- AI Video — Scene Inconsistency Between Cuts
- AI Video — Prompt Keyword Ignored Mid-Clip
- AI Video — Over-Motion / Too Much Movement
External references:
- Runway: Inpainting (official help docs) — masking and regenerating a region of a clip.
- Kling: Multi-Reference and Inpainting guide — redraw-intensity behavior for local edits.
- Kling 3.0 prompt guide — the “anchor hands to an object” technique for stable grip geometry.
Tags: #ai-video #Troubleshooting #Video generation #hands #motion-artifacts #anatomy