Choosing a reference for a live transformation

The reference is an input to the model, not a promise of the final result. Start with an image you own or have permission to use, then check a short generated session before broadcasting.

Face references

Use one clear face with even lighting and a front or slight three-quarter angle. Avoid heavy occlusion, very small faces and multiple people. In the camera preview, keep one person visible and make a few slow head turns. The reference presets in the library are useful inputs to explore; they are not real generation samples.

Outfit references

Pick an image that shows the garment clearly. Frame the matching part of your body in the camera. A coat reference will be hard to interpret if the camera only shows your face. Start with a simple background, then check sleeves and edges during movement. Visual try-on does not establish garment size or physical fit.

Style references

Choose a reference with a consistent visual treatment, such as watercolor or print. Highly mixed styles can make the result less predictable. Use the style mode and inspect both your subject and the surrounding scene.

Camera access and a stable input

Open the site over HTTPS (or localhost for local development), then allow camera access in the browser's site settings. Also allow the browser to use the camera in your operating system's privacy settings. After changing a permission, reload the page and open the camera again. This preview does not need microphone permission.

If no camera appears, reconnect it and select the intended device. If it is busy, close other apps using the physical camera and retry. Stop generation before switching cameras. Use Device check to inspect the camera's reported resolution and frame rate; these describe the input, not the quality or speed of generated output.

Flicker, glasses and a changing face

• Compare the camera preview with the generated result. If the preview flickers too, try steady front lighting, move away from a bright window and keep the camera still.
• Start facing forward and turn slowly. Fast motion, a full side profile, hands across the face or hair over the eyes can make the result less stable.
• With glasses, adjust the light angle to reduce reflections and choose a reference with clearly visible eyes. Check the result both still and while turning your head.
• If the generated face changes identity or its edges jump, return to a clear single-face reference, reduce occlusion and change one thing at a time. A clearer input can help diagnosis, but cannot guarantee a fix for a model limitation.

Usage Guide：Read the usage guide and compare the moving preview while you turn your head.

Save only what you need

You can keep up to 60 images and preset favorites in your private library. Saved uploads are normalized and encrypted. Deleting one clears the current stored image and thumbnail; copies already sent to a provider require a separate request through Privacy & reports.

The current service accepts JPEG, PNG and WebP. Local preparation rejects very small images; it cannot certify face identity, image ownership or model suitability. Real provider output and latency require a configured service and hands-on verification.