A still photo has no motion in it. Not hidden motion, not latent motion you can extract — nothing. One frame, no time axis, no information about what happened a tenth of a second before or after. So when a video comes back with the person in that photo blinking, turning their head, and shifting their weight, something has to be invented.
Understanding what gets invented, and where the invention is anchored, explains almost every quality question people have about AI video: why the face holds together while the jacket changes colour, why the hands go wrong, and why a ten-second clip is a fundamentally different task from a thirty-second one.
What animating a photo is not
It isn't pixel displacement. Nobody is warping your image around a mesh, sliding the eyes down and the mouth open. That technique exists — it's how older face-swap apps worked — and its signature failure is a rubbery, sliding quality where features drift across the skin.
Modern image-to-video generation doesn't transform your photo. It generates new frames, and uses your photo as a constraint on what those frames are allowed to contain. The distinction matters because it explains the failure modes: a warp can't invent a background, but a generator can, and sometimes does.
Step one: the photo becomes a description
The first thing that happens to your uploaded image is that it gets encoded into an internal representation — a dense numeric summary of what's in the frame rather than a grid of pixels. Layout, shapes, tones, the rough geometry of a face, the colour of the light.
This is a lossy step, and it's lossy in a specific direction: semantic content survives, fine detail doesn't. "A person with dark hair facing the camera in a dim room" survives comfortably. The exact pore pattern on their cheek does not. This is why over-filtered photos produce worse results than raw ones — a beauty filter removes the micro-texture that the encoder was going to use to reconstruct skin, leaving nothing but a smooth gradient, and the generator fills the gap with something generic and waxy.
Step two: frames are generated, and the photo votes
Each frame starts as noise and gets progressively refined. At every refinement step, the model is pulled toward two things at once: the text description of the scene, and the reference image.
The reference is why the result resembles your photos at all. It's also why the resemblance is uneven. The constraint is strongest where the reference is most informative — a well-lit frontal face is packed with distinctive structure, so the model has a lot to hold onto. A hand partly out of frame, or an arm in shadow, carries almost no information, so the model is essentially free to invent, and it invents badly because hands are hard.
Step three: consistency is between frames, not within them
Here's the part most explanations skip. A single convincing frame is not hard. A hundred frames that all show the same person are hard, because nothing in the process inherently links frame 40 to frame 39.
What links them is a second constraint: each new frame is also conditioned on the frames just before it. The model is asked to continue a motion, not to render an independent picture. This is why clips hold together at all, and it's also why the failures cluster at moments of change.
Three things reliably strain that link:
- Occlusion. When an arm passes in front of the torso, the model has to remember what was behind it. It often doesn't, and reconstructs a slightly different shirt.
- Large motion. A fast turn means most of the frame is new information with no prior frame to condition on.
- A cut. A hard cut resets the conditioning. Everything after it is generated from the text description and the reference image again, which is why clothing, lighting, and even face shape can shift across a cut while staying perfectly stable within a shot.
Why faces hold and hands drift
The asymmetry is not mysterious. Information density decides it.
A frontal face at a decent resolution gives the model hundreds of stable landmarks — eye spacing, brow line, nose shape, the specific way the mouth sits. Those landmarks are re-checked against the reference on every frame, so they stay put.
A hand at ten seconds into a clip may be small, angled, partly shadowed, and moving. Almost none of that is in the reference photo. The model has to synthesise finger structure from scratch, and fingers are one of the hardest things in visual generation because a small positional error turns a hand into something with the wrong number of joints.
The practical takeaway is not "AI can't do hands." It's that a hand which is visible, well-lit, and roughly in the same pose as the reference photo will look fine, and a hand that gets invented from nothing probably won't.
Why the eyes are the tell
Humans are calibrated to read eyes with absurd precision. We can detect a gaze that's slightly off, a pupil that's the wrong size, or a reflection that doesn't match the lighting, and we do it pre-consciously.
In a generated video, the eyes are doing two jobs at once: they have to match the reference photo's specific eye shape, and they have to move like eyes — micro-saccades, blinks at plausible intervals, a gaze that lands somewhere rather than floating. When those two jobs conflict, the eyes are the first place it shows, and the result reads as "off" to viewers who can't say why.
This is also why a frontal, sharp, evenly-lit photo outperforms a more flattering angled one. The eyes are the highest-information feature in the frame, and a clear view of them is the single best thing you can give the model.
Why ten seconds is a different problem from thirty
Consistency degrades over time the way any iterative process accumulates error. Frame 300 is conditioned on a chain of 299 predecessors, each of which introduced a small deviation.
In practice, this means the first few seconds of a clip are usually the strongest, and drift compounds from there. It's the reason short clips feel tight and longer ones feel like they're dissolving. It's also why the same model can produce a flawless ten seconds and a mushy thirty — nothing changed except how long the chain got.
Knowing this changes how you'd plan a project. Instead of one thirty-second video, three ten-second shots cut together will look better than a single long generation, because each shot restarts the chain with a fresh, strong reference.
What this means for the photo you upload
All of the above collapses into a short list:
- Frontal beats artistic. A clear view of both eyes is worth more than a flattering angle.
- Raw beats filtered. Skin texture is signal. Smoothing deletes it.
- Well-lit beats moody. Shadow removes information the model would otherwise use.
- Sharp beats everything. A soft or blurred face is a face the model has to guess at.
- One person per photo. A couple photo has to be split into two characters, and the split is where features bleed.
Frequently asked questions
Does the model copy my photo or generate a new person?
It generates someone who matches the photo's structure closely. At a ten-second clip viewed at normal speed, the resemblance is convincing. Paused and zoomed, you'll find differences — the model is reconstructing, not tracing.
Why do the clothes change but the face doesn't?
Clothing carries less identity information than a face, so the model has less reason to hold it consistent, and occlusion during movement gives it an excuse to re-invent. The face is the anchor; the wardrobe is negotiable.
Can I fix a bad result by describing the problem in the prompt?
Sometimes, if the problem is scene-level (wrong light, wrong location, too much motion). Prompting won't fix a photo-level problem like a turned face or a heavy filter — that has to be solved by uploading a different photo.
Does resolution of my uploaded photo matter?
Yes, meaningfully. 256 × 256 is the floor, not the target. A larger, sharper source image gives the encoder more structure to preserve, which shows up most in the eyes and around the mouth.
Why does my video look great for three seconds and then drift?
Drift is accumulated error along the frame chain. It's inherent to the approach, and the main practical workaround is shorter shots rather than longer ones.
The part that's actually under your control
Everything expensive about this process — the retries, the disappointing renders, the prompt rewriting — traces back to inputs chosen in the first two minutes. The model is doing something genuinely difficult with whatever it's given; the leverage is in what you give it.
If you want to see the motion behaviour before committing a photo, the sample on the generator page shows the shot rhythm and how the model handles light. What it won't show is your specific result, because it's the official demo — the honest version of "preview" in a category where most previews are marketing.