Choosing photos that actually work

Why the input photo decides the output

The generator does not invent a person. It works from the face and body in your picture and reconstructs them inside a new scene. Everything the model cannot see clearly, it has to guess — and guesses are where likeness drifts, faces soften and details wander.

That means your photo is the raw material, not a reference. A great photo produces a great clip; a poor photo caps the result no matter which trend you choose. Improving the input is the highest-leverage thing you can do, and it costs nothing.

Resolution and sharpness

Feed the model enough pixels to work with. Under about half a megapixel, or roughly 256 pixels on the short edge, facial structure starts to break down and the output looks painted rather than photographic.

Sharpness matters more than raw size. A 2-megapixel photo with a crisp face beats a 12-megapixel photo that is soft, motion-blurred or out of focus. Avoid photos taken at night without enough light, and avoid digital zoom, which enlarges noise instead of detail.

• Aim for the face to occupy a solid portion of the frame — a head-and-shoulders portrait is ideal.
• Prefer a well-lit, in-focus photo over a huge but blurry one.
• Do not crop down to a tiny face extracted from a group photo.

Framing: head-and-shoulders wins

For most trends, the sweet spot is a head-and-shoulders shot in which the face takes up a meaningful part of the image. This gives the model clear information about the features it needs to reproduce.

Very wide shots — full-body photos taken from across a room, or faces lost in a crowd — leave too few pixels on the face. Meanwhile, extreme close-ups that cut off the top of the head or the chin can confuse the reconstruction. Give the model the whole head with a little breathing room.

Full-body framing is only needed when the trend itself performs a full-body action, such as a dance. Match the framing to what the trend has to do.

One person per photo (for duet templates)

Duet-style trends place two people side by side. Upload one photo per person rather than a single photo containing both subjects. The model draws each subject from their own image, which keeps both likenesses sharp.

It also helps if the two photos are lit and framed similarly. A bright outdoor selfie next to a dim indoor shot can look mismatched once they stand in the same scene. Matching the lighting and the general distance from camera keeps the pair believable.

Lighting: soft and even beats dramatic

Even, diffuse light reveals facial structure. Harsh side light, strong shadows across half the face, or a bright window behind the subject all hide the details the model needs. Overcast daylight, a north-facing window, or a soft indoor lamp are ideal.

A useful test: if you can clearly see both eyes and the shape of the nose and cheekbones without deep shadow, the lighting is good enough. If half the face is lost to shadow, find a softer spot before uploading.

Backgrounds: simple beats busy

A calm, uncluttered background lets the model separate you from the photo cleanly. Busy scenes — a crowded street, a shelf of objects, strong patterns — can bleed into the reconstruction of the subject.

You do not need a studio backdrop. A plain wall, a blurred outdoor scene or a simple interior all work well. The goal is to make it obvious where the person ends and the background begins.

Filters, screenshots and compression

Heavy beauty filters smooth away the very texture the model uses to recognise a face. The result often looks like a different, generic person. A normal, lightly processed photo reproduces likeness far more reliably than an edited one.

Screenshots and heavily re-shared images are also risky, because they have been compressed again and again. Each round of compression loses detail. Where possible, use the original photo straight from the camera roll rather than a screenshot of a photo.

• Avoid strong beauty filters and skin-smoothing apps.
• Avoid screenshots of photos — export or send the original instead.
• Avoid photos that have been cropped and re-saved many times.

Run the free photo check first

Before you spend credits, put your photo through the free readiness check. It scores the image on resolution, framing, lighting and subject clarity, and tells you whether it is ready, worth improving, or likely to fail.

The check runs in your browser and returns an instant result without requiring a generation. A thirty-second pass here is the cheapest way to avoid a disappointing video later.

A photo that passes the check is not a guarantee — generative video is never perfectly predictable — but it removes the most common causes of weak results.

How many photos should I upload?

One per person is the rule. Single-subject trends take one photo; duet trends take two — one photo for each person who should appear on screen. Uploading a single photo that already contains two people into a duet slot tends to produce a merged, blurred result rather than a clean pair.

If you have several candidates for the same subject, decide which is cleanest before you start: the sharpest, best-lit, least-filtered option. You can always try a second photo afterwards, but starting from your strongest image gives the most useful first result and the clearest signal about what to fix next.

Rescuing a photo you love

Sometimes the only photo you have of someone is not ideal — a little soft, slightly dark, or an old screenshot. That does not make it unusable, but it does lower the ceiling on the result, so it is worth giving the model every advantage you can.

• Crop to the face and shoulders so more pixels land on the features that matter.
• Go back to the original file where you can — every re-save and screenshot loses detail.
• Do not over-sharpen or heavily upscale; that adds artefacts, not real detail.
• If only one eye is visible, or the face is turned far to one side, prefer a more frontal shot.

A rescued photo can still produce a good clip — just expect the likeness to be a little looser than an ideal input would give.

Before you commit: a thirty-second pass

Run a quick mental checklist against every photo before it goes in. Is the face large and sharp? Is the lighting even, with both eyes clearly visible? Is one person the clear subject? Is the background simple? Is the image the original, not a screenshot?

If every answer is yes, upload it with confidence. If one answer is no, that is the single thing to fix — and fixing it is usually faster than re-generating a clip that was never going to look right.

Portrait or full-body — what each trend wants

The framing that works best depends on what the trend has to do on screen. Trends that keep the subject close — a talking-head style, a subtle head turn — want a tight portrait where the face fills a good part of the frame. Trends that perform a full-body action — a walk, a dance — need the whole body visible, which means a wider shot.

Decide the trend before you choose the photo, then match the two. Choosing a photo first and forcing it into a trend it cannot support is the source of a lot of weak first attempts. If a photo only works for one kind of trend, that is useful information, not a limitation.

When in doubt, a head-and-shoulders photo is the more versatile choice — it suits more trends than a full-body shot does.