Most disappointing talking-photo results come from the input, not the model: a face that is too small, turned away, blurred by motion or lost in shadow. Checking first saves you a wasted render and shows exactly what to change before you add a voice.
What the talking photo checker looks at
Every photo goes through nine checks. The first three are hard limits of OmniHuman 2; the rest are quality signals that decide how natural the motion looks.
- File: JPG, PNG, WebP and other common formats up to 30 MB.
- Resolution: at least 300 px on the shorter side, no more than 6000 px on the longer side; 720 px or more is recommended.
- Aspect ratio: anything from 2:5 (very tall) to 5:2 (very wide), so phone portraits, squares and widescreen frames all qualify.
- Face: whether a face is found at all, and how many, using an on-device face detector.
- Face size: how much of the frame the face fills, and how many pixels the mouth and eyes get.
- Framing: whether the head is cut off by the edge of the picture.
- Head angle: whether the person roughly faces the camera and the head is upright.
- Sharpness: a blur measure computed on the face itself, not on the background.
- Lighting: whether the face is too dark, blown out or so flat that features wash away.
How to read your results
A green tick means the check passed. An amber tip means the photo will work, but changing that detail is likely to improve the video. A red mark on file, resolution or aspect ratio means the studio will refuse the file until you fix it, usually by exporting a JPG, resizing or cropping.
If no face is found, the photo can still be tried: stylised art, statues and animals sometimes animate well, but results are less predictable than with a clear human portrait. The face box drawn on your preview shows exactly what the detector saw, which helps when a group photo or a busy background confuses it.
Quick fixes for common problems
Almost every warning can be solved in a minute with your phone’s built-in editor or any free photo app:
- 1Face too smallCrop in so the head and shoulders fill most of the frame. Cropping a large original keeps more detail than enlarging a small one.
- 2Head turned or tiltedPick a shot where both eyes are visible and the nose points roughly at the camera, or straighten a tilted scan with the rotate tool.
- 3Blurry or softUse the original file instead of a screenshot or a messaging-app copy, which are heavily compressed. Avoid shots taken while moving.
- 4Too dark or too brightRaise shadows or lower highlights until skin shows texture. Soft daylight from a window is the most forgiving light for animation.
Private by design: nothing is uploaded
The checker downloads a small face-detection model (Google’s open-source MediaPipe BlazeFace) and runs it inside your browser with WebAssembly. Measurements like blur and brightness are calculated on a canvas on your device. Your photo is not sent to our servers or anyone else’s, and closing the tab discards it.
Because there is no server work, the tool is free to use as often as you like, with no sign-up and no watermark. It works on recent versions of Chrome, Edge, Safari and Firefox on desktop and mobile.
From checked photo to talking video
When the checker says your portrait is ready, the Make it talk button carries the same photo into the OmniHuman 2.0 studio, so you don’t have to upload it again. Sign in with Google, add a voice recording or audio file between 4 seconds and 5 minutes, and preview the result before choosing a plan.
Planning a group scene, a song or a product spokesperson? The guides below explain how to prepare audio and prompts for each case.
FAQ
Is the talking photo checker really free?
Yes. It runs on your own device, so there is no cost to us and no limit, account or watermark for you.
Is my photo uploaded anywhere?
No. Face detection and all measurements happen in your browser. It is only uploaded if you continue into the studio and create a video with it.
Does a pass guarantee a perfect video?
No tool can promise that, but a pass removes the most common causes of stiff or smeared lips. Clean audio matters just as much.
Why does it say no face was found?
The detector looks for human faces roughly facing the camera. Very small faces, profiles, heavy shadows and cartoons can be missed.
Can I check a group photo?
Yes. Every face it finds gets a box. OmniHuman 2 can animate several people; describe who speaks in the prompt.
Does it work for other AI talking-photo apps?
The quality checks apply to any lip-sync model. File and size limits are specific to OmniHuman 2.
Which formats can I check?
JPG, PNG, WebP, GIF and BMP work everywhere. HEIC opens in Safari; elsewhere, export it as JPG first.