Why AI Video Distorts Hands and Faces on VdoBloom (And the Five Fixes That Work)

Hands and faces distort in AI video because they are the smallest, most detailed and fastest-moving parts of a frame — the model spends the least pixel budget exactly where your eye looks hardest — and the fix is almost never a cleverer prompt on its own: it is a cleaner source photo, a higher resolution tier, a shorter clip, and a model that was built to hold identity. On VdoBloom the whole repair loop costs a handful of credits, because an image edit runs at 5–7 credits and a 5-second 720p test clip runs at 32–100 credits depending on the model you test with.

What is actually going wrong

Three separate problems get lumped together as “bad hands.” They have different causes and different fixes, so separate them before you burn credits re-rolling the same clip.

Resolution starvation. A hand in a wide shot at 480p occupies maybe 20 by 20 pixels. There is no room in that grid for five distinguishable fingers, so the model paints a plausible blob. Nothing in your prompt changes the arithmetic. This is why the same prompt that produces mush at 480p produces a clean hand at 1080p.

Motion ambiguity. Fingers occlude each other constantly. When a hand rotates, the model invents what was hidden and must re-invent it consistently frame after frame. The longer the clip, the more chances it has to guess differently — which is why finger errors usually appear in the second half of a 10-second clip.

Identity drift. Faces do not usually explode; they slide. Frame one looks like your subject, frame ninety looks like their cousin. This is a different failure from hands and it is largely a model-choice and prompt-anchoring problem, not a resolution problem.

The five failure modes and the fix that actually works

What you see	Real cause	The fix	Typical cost on VdoBloom
Six fingers, fused fingers, melting knuckles	Resolution starvation in a wide shot	Move up a resolution tier, or reframe closer so the hand is bigger in frame	Wan 3.0 at 5s: 16 credits at 480p vs 64 at 1080p
Hand fine at the start, wrong by the end	Motion ambiguity across a long clip	Cut duration. Generate 5s, not 15s, then extend or cut	Kling 3.0 Omni 720p: 40 credits at 5s, 120 at 15s
Face slowly becomes a different person	Identity drift — weak identity anchoring	Use image-to-video with a sharp face-forward photo and an identity-preserving prompt	Same cost as any image-to-video run
Face is smeared or plastic from frame one	The source photo was already soft, small or over-filtered	Fix the still first: edit or upscale, then animate	Nano Banana Edit 5 credits; Flux Kontext Pro 7; Seedream 4.5 Edit 7
Teeth, eyes or jewellery flicker	High-frequency detail the codec and the model both fight over	Render at 1080p and upscale the finished clip, rather than trying to rescue a 480p render	Topaz video upscale: 3.2 credits per second at 2x, 5.6 at 4x

Fix 1: repair the still before you animate it

A video model cannot add detail that was never in your input frame — it only propagates what is there. If the face in your photo is 300 pixels wide, soft and smoothed by a phone beauty filter, every output frame inherits that.

Run the photo through an image edit first. Nano Banana Edit costs 5 credits per image, Flux Kontext Pro 7, Seedream 4.5 Edit 7 — three cleanup attempts cost less than one 5-second video. The AI image upscaler handles resolution; image editing handles hands that were already wrong in the photograph. Fixing a sixth finger in a still is a 5-credit problem. Fixing it across 150 frames is not fixable at all.

If your photo shows a real person, you must have that person’s consent before you animate them — that applies to every photo-motion effect on the platform, not just the obvious ones.

Fix 2: choose the resolution tier before the duration

Most people pick 10 seconds at 480p because it feels economical. It is the worst of both worlds: you paid for the frames where hands go wrong and starved the pixels that would have kept them right. Invert it — generate 5 seconds at 1080p, then decide whether you need length.

Model and tier (5 seconds)	Credits	Cost at $0.05 per credit	When to use it
Wan 3.0, 480p	16	$0.80	Blocking a shot. Never for close-up hands
Wan 3.0, 720p	32	$1.60	Cheap iteration where a face is mid-frame
Wan 3.0, 1080p	64	$3.20	Best value 1080p for detail-critical clips
Kling 3.0 Omni, 720p	40	$2.00	Strong identity hold with native audio
Kling 3.0 Omni, 1080p	54	$2.70	The default for faces in close-up
Seedance 2.5, 720p	100	$5.00	Complex motion and camera work
Seedance 2.5, 1080p	180	$9.00	Final hero shot only

Read that table as strategy: three 1080p Wan 3.0 tests cost less than one 1080p Seedance 2.5 run. Prove the shot cheaply at the good resolution, then spend once.

Fix 3: anchor identity in the prompt, do not describe it

The prompts VdoBloom ships with its photo-motion effects all contain a version of the same clause, and it is there for exactly this reason: Preserve facial features, body proportions, clothing, lighting, shadows and background exactly as in the original image. That instruction does more for face stability than any amount of adjective stacking.

What does not help: describing the face you already uploaded. Writing “beautiful symmetrical face, perfect hands, five fingers” pushes the model toward a generic idealised face and away from the specific one in your photo. Negative-style phrasing is worse — the video models in the VdoBloom picker take a single positive prompt, so a phrase like “no extra fingers” is read as content, not as a prohibition.

Fix 4: compose hands out of the danger zone

Hands are reliable when they are large, slow and unoccluded, and unreliable when they are small, fast and crossing the body. If the shot does not need the hands, frame them out or give them something to rest on. Motion templates that keep arms in a clear, repeating arc — a fashion walk or a hair flip — hold up far better than anything that asks for fine manipulation, because the model never has to solve finger occlusion.

Fix 5: pick a model that holds people

Model choice is real and it is measurable. On the VdoBloom catalogue, Kling 3.0 Omni image-to-video is the strongest default for human subjects: it runs 3–15 seconds at 720p, 1080p or 4K, and in image-to-video mode the aspect ratio follows your input image, so the face is never letterboxed or cropped into fewer pixels. Wan 3.0 is the value pick for iteration, at 480p, 720p or 1080p and any duration from 2 to 30 seconds. Seedance 2.5 earns its price on complex camera moves, not static portraits.

A worked repair, start to finish

1. Upload the photo. Face is soft. Run one Nano Banana Edit pass to sharpen and clean the hand — 5 credits.
2. Test the motion on Wan 3.0 at 720p for 5 seconds — 32 credits. Hands hold, face drifts slightly at the end.
3. Re-run at 5 seconds on Kling 3.0 Omni at 1080p — 54 credits. Face holds.
4. Total spend: 91 credits, about $4.55, with two throwaway attempts included.

Compare that to generating straight to 15 seconds of 1080p Seedance 2.5 and discovering the hand fails at second eleven. New accounts get 10 free credits with no card, enough to run the image-edit step and find out whether your source photo was the real problem. Tier costs are on the pricing page; one-time credit packs start at $2.49 and never expire.

Frequently asked questions

Does a longer prompt fix bad hands?

No. Prompt length has no effect on the pixel budget a hand receives. Resolution, framing and clip length do. Use the prompt to pin identity and describe motion, and solve the hand problem with the resolution tier and the shot design.

Will upscaling a finished video fix distorted fingers?

It will sharpen them, not correct them. Topaz upscale on VdoBloom bills per second of source video — 3.2 credits per second at 2x and 5.6 at 4x — and it makes a good clip crisper. An upscaled six-fingered hand is a sharper six-fingered hand.

Why does the face change halfway through the clip?

Identity drift accumulates over frames. Shorten the clip, use image-to-video rather than text-to-video so there is a real reference to hold onto, and include the preserve-identity clause. If it still drifts, move to a model with stronger identity hold such as Kling 3.0 Omni.

Is 480p ever the right choice for a clip with people in it?

Yes, for testing motion and timing, where you only need to know whether the movement reads correctly. It is 16 credits on Wan 3.0 for 5 seconds. It is never the right choice for a delivered clip with visible hands or a face in close-up.