Photo to dance video: what the AI keeps and invents

A photo-to-dance render in this studio is a controlled guess. The model takes one still image, reads the subject in that frame, and builds a short video around it. Some details survive because they are visible in the file. Others are invented because the model has to create motion that was never there. Knowing which is which sets realistic expectations before you spend any credits.

What carries over from the still

The photo constrains the render most strongly in the face and upper body. The studio's upload guidance asks for a sharp, well-lit picture of one person or animal, taken from the front or a three-quarter angle, with the face and body visible. That framing gives the model the most detail to hold onto. In practice, these details travel from still to video:

| Detail | How well it usually travels | What makes it less reliable | | Face shape and features | Strong when the face is sharp and unobscured | Profile view, shadow, motion blur, very small face | | Hair length and color | Strong when the whole hair outline is visible | Cropped crown, hair over the face | | Body shape and upper body proportion | Moderate, read mostly from the visible torso | Tight face-only crop, wide shot with small subject | | Clothing color and simple style | Moderate | Busy prints, low resolution, harsh shadows | | Background environment | Weak, often rebuilt | Cluttered scenes, object detail |

The model reads the photo as a starting point, not as a pixel copy. The closer a feature is to the face and the clearer it appears, the more it persists. A soft or distant face gives the model less to preserve, so the result will drift toward a generic face that has the right general shape but less of your specific detail.

What the model invents

The model does not recover hidden motion from a still. It invents new motion using the chosen look as the direction. That means several parts of the clip are generated rather than carried over:

• The dance itself: arm swings, weight shifts, hip movement, and timing all come from the model. No reference video is supplied, and there is no motion transfer from another clip.
• Background movement: the plane behind the subject may slide, simplify, or change lighting.
• Cloth and hair motion: the model adds folds, sway, and bounce that were not in the photo.
• Facial expression changes: the face may blink, smile, or shift gaze, and small asymmetries can appear.
• Camera framing: the model decides how the generated body stays in frame, including zoom or position changes.

Because these elements are generated, two renders from the same photo and the same prompt will not match. Rerunning can produce a slightly different dance, different hand positions, or a different background. This is not a bug; it is how the generation works.

Common artifacts in generated dance clips

Artifacts are most visible in hands, boundaries between people and background, and anything that moves quickly. The table below lists the usual failures and the source photo or prompt changes that reduce them.

| Artifact | What it looks like | Frequent cause | Change to make | | Warped hands or extra fingers | Hands bend oddly, merge with clothing, or show extra digits | Hands partly hidden, tight crop, fast motion | Use a photo with hands visible or frame from chest up; ask for one clear movement | | Limb sliding or merging | Arms or legs appear to pass through each other or combine | Two subjects, body parts overlapping, low source detail | Crop to one subject, keep limbs separated | | Face drift during a turn | The face changes slightly when the subject turns | Profile or shadowed source, heavy compression | Use a front-facing or three-quarter photo with even light | | Clothing pattern swimming | Stripes or prints move and reform | Busy pattern | Choose solid clothing or describe the garment | | Background warping | Objects behind the subject bend or disappear | Cluttered or detailed background | Use a plain background or state the setting in the prompt | | Flicker between frames | The whole frame pulses or the subject jumps | Very low resolution source or strong motion request | Use a sharper source, shorten the clip, or simplify the prompt |

Brief hand distortion can still appear in otherwise good renders. If the clip reads well at normal speed and the face stays consistent, a single odd frame may not be worth another render.

Length and resolution trade-offs

The studio offers 5, 8, 10, and 12 second clips and 480p, 720p, and 1080p output, with 9:16, 16:9, and 1:1 frames. The default is 720p at 12 seconds, which costs 26 credits. A 480p 12-second render costs 12 credits, and a 1080p 12-second render costs 64 credits. Costs scale linearly from those 12-second rates and round up to a whole credit when you change length.

These settings trade cost against visible detail and motion risk.

• 480p is the cheapest way to test a photo. It hides some fine artifacts but also hides fine facial detail. Use it for a first pass.
• 720p is the standard social media balance. It shows enough detail to judge the face and the dance, and it matches the default price point.
• 1080p gives the sharpest output but makes artifacts more obvious. If the source photo is soft or the pose is ambiguous, higher resolution does not fix that problem.

Longer clips cost more and give the model more time to drift. A 5-second clip at 480p is the lowest-risk test for a new photo because there are fewer generated frames in which hands, hair, and background can move. If a photo passes that test, a 12-second 720p render is more likely to be worth the credits. If a photo fails at 5 seconds, it will usually fail worse at 12 seconds.

Each account gets one free preview render at 480p for 5 seconds while the daily preview budget lasts. Email sign-up starts with a zero credit balance. Google sign-in accounts can claim 10 free credits per day for three days, separate from that preview. There is no unlimited free-forever plan. On the pricing side, the $9.90 single pack contains 10 credits, so it does not cover one standard 26-credit render. The smallest one-time pack that covers a standard 720p 12-second video is the $19.90 Duo Pack with 30 credits.

When a result is as good as it gets

A render is likely as good as it gets when the following checklist is met:

1. The face stays recognizable for the full clip.
2. Arms and hands track the dance without repeated merging or disappearance.
3. The body stays inside the frame for the entire duration.
4. Clothing reads as the same garment rather than a shifting pattern.
5. The background stays stable enough not to distract.

If all five are true at 720p and 12 seconds, further attempts at 1080p or a longer length will often produce marginal gains instead of large improvements. The model may still show a short hand warp or a background wobble. Those are expected in generated dance video.

Do not chase a perfect version of a clip that already meets the checklist. The randomness in generation means rerunning can fix one artifact and create another. When the clip works for the intended platform and the subject is clear, that is the point to stop.

For troubleshooting persistent problems, the troubleshooting guide covers failed renders and retry decisions. Credit costs for every resolution and plan are listed on the pricing page. You can also open the studio and run the free preview before committing credits.

FAQ

Does the photo control the dance or just the face?

Both, but in different amounts. The photo constrains the face, hair, upper body, and clothing most strongly. The dance motion, background, and small expression changes are generated. The photo cannot supply choreography, so the model invents that part.

How much detail should I expect in the clothes?

Simple, solid-colored clothing carries over more faithfully than busy prints. The model may still add folds and movement that were not in the still. If a garment is central to the look, describe it in the prompt and keep the pattern plain.

Is 1080p always better?

No. 1080p is sharper, but it also makes artifacts more visible. If the source photo is soft or the face is small, 1080p will not add missing detail. Start at 480p for a test, then move to 720p for the final social clip, and use 1080p only when the photo already has enough detail.

Why do hands look wrong?

Generated video models often struggle with hands, especially when fingers are partly hidden, blurred, or moving quickly. A photo with visible hands and a shorter clip or simpler prompt can reduce this. Brief hand distortion in an otherwise good clip is common.

When should I stop rerendering?

Stop when the five-point checklist in this guide is met. If the face stays consistent, the body stays in frame, the clothing reads clearly, and the background is stable, extra rerenders are unlikely to give a large improvement. Each rerender uses credits and can introduce a new artifact.