People ask why do ai videos look weird at the exact moment a clip stops being impressive and starts being unsettling. The answer is rarely "the model is bad". Six specific causes account for almost every bad result, and five of them are fixed before you touch a setting.
The causes are ordered from most to least common, based on what goes wrong with two-photo renders.
One: the guidance was set too loose
Every image-to-video system balances two goals: stay close to your photo, and produce visible motion. Turn up motion and you turn down fidelity. The face is the first thing to slide.
The mechanism is a denoising trade-off formalised years before these products existed, in the classifier-free guidance paper. Loose guidance means the model trusts its own prior more than your frame, and its prior is an average of millions of faces — which is exactly what a drifting face looks like.
The fix: reduce the camera move before you reduce anything else.
Two: the face was too small in the source
This is the most common non-setting cause. A full-body photo encodes a face around 100 pixels wide. At that size the encoder keeps "human face, roughly this shape" and discards the features that make it yours.
The fix: crop to head and shoulders. If the crop is too soft to use, the photo was never a candidate.
Three: the background is being invented
The model does not preserve your background; it regenerates one consistent with your description. A patterned wall becomes a plausible wall. A crowded street becomes a plausible street with different people.
The scene is invented. Your faces are the constraint.
That sentence is the single most useful thing to hold onto, and it comes straight from how the architecture works (arXiv:2311.15127).
The fix: pick photos with plain backgrounds when you care about the setting. Or lean in and choose a template whose scene you actually want.
Four: everything is too smooth
Compression removes high-frequency detail. Skin texture, fabric weave and hair strands are all high-frequency. What survives is a smooth surface that reads as plastic.
This is the same trade-off the latent diffusion authors describe as "decomposing the image formation process into a sequential application of denoising autoencoders" (arXiv:2112.10752) — the compression is the method, not a bug.
The fix: start from a photo with visible texture and hard light. A phone flash portrait has more usable high-frequency detail than a soft studio shot.
Five: hands, teeth and ears
These three regions are where the model has the least training signal for temporal consistency. Fingers merge, teeth blur into a smear, ears change shape between frames.
The fix: keep hands out of frame, keep mouths closed, and avoid earrings that draw attention to the ear line.
Six: the clip is too long
Every additional second is another chance for small errors to accumulate. A 4-second clip with one error reads as a stylistic choice. A 10-second clip with the same error reads as broken.
The fix: cut to the shortest clip that carries the idea. Four to six seconds is the sweet spot for a two-photo render.
Why do ai videos look weird: the diagnostic table
| What you see | Most likely cause | First thing to change |
|---|---|---|
| Face morphs into a stranger | Guidance too loose | Remove camera move |
| Waxy, plastic skin | Over-smoothing | Swap to a harder-lit photo |
| Background looks like a different room | Scene invention | Choose a different template |
| Fingers merge or vanish | Hand region | Reframe to exclude hands |
| Teeth smear | Open mouth | Use a closed-mouth photo |
| Errors appear after ~6 seconds | Duration | Shorten the clip |
Work down the table in order. Changing the template first is the most common mistake, because it resets everything and teaches you nothing.
What "weird" usually means, precisely
The uncanny feeling has a specific source: the clip is photorealistic in texture but wrong in behaviour. Eyes that do not track, a head that turns without the neck following, a smile that arrives a beat late. Photorealism raises the expectation of correct physics, and the gap between the two is what people describe as weird.
Which is why horror is such a good fit for this technology. A zombie is allowed to move wrongly. The same clip that looks broken in a romance video looks deliberate in a haunted one.
Frequently asked questions
Why do ai videos look weird even when the photo is good?
Usually duration or motion. A good photo with a large camera move will still drift, because the motion setting overrides fidelity. Shorten the clip and reduce the move before changing anything else.
Is a higher resolution source always better?
No. Resolution does not restore detail the encoder discarded, and a soft high-resolution photo performs worse than a sharp low-resolution one. Sharpness at the face is what matters.
Can a prompt fix a drifting face?
No. Prompts control scene and mood, not identity. If the face drifts, the input photo or the motion setting is the cause.
Do all tools produce the same artefacts?
The failure modes are shared because the architecture is shared; the severity differs by product. That is also why the same fix tends to work across tools — and why do ai videos look weird is a question about physics, not about brands.
Render one clip at the shortest duration your template allows, then render the same photo again with a bigger camera move. Compare the two — the difference is the entire answer to why do ai videos look weird.