SadTalker Input Checker
Check a portrait and narration before spending time on a render. Get file details and practical warnings, then review the face yourself.
Selected files are processed in your browser and are not uploaded. The example buttons load media already on this website. These checks are heuristics, not a prediction of animation quality.
Choose either file or check both. Your selected files stay on this device.
1. Check an image
PNG, JPEG or WebP. Up to 20 MB and 24 megapixels.
2. Check audio
WAV, MP3, M4A, OGG or FLAC, if your browser decodes it. Up to 30 MB and 3 minutes.
How to use the results
- Select an image. Review its dimensions and preview, then complete the face checklist.
- Select narration. Check duration and level warnings, and listen to the file for unwanted sounds.
- Resolve input problems first. Use the Settings Finder to choose starting settings and the installation guide to prepare a command.
- Render a short excerpt and inspect mouth timing, likeness, eyes and face boundaries.
What the checks measure
Image: dimensions refer to the decoded image, not the face size. Sharpness is the variance of a four-neighbour grayscale Laplacian after resizing the longest side to at most 512 px. Values below 80 trigger a low-edge-detail warning. This threshold is a starting heuristic, not calibrated against SadTalker output quality. Texture, contrast, noise and background can change the result; a sharp background can hide a soft face.
Audio: peak and RMS levels use decoded samples from all channels without mixing them together. dBFS is relative to digital full scale. A warning appears below −30 dBFS RMS, above 0.1% near-full-scale samples (absolute value at least 0.999), or when more than 40% of 50 ms windows have RMS below −40 dBFS. These warnings describe levels and pauses, not a measured noise floor or speech-to-noise ratio.
Browser limits: decoding support varies by codec and browser. Audio sample rate may be resampled to the browser’s audio context rate and is not necessarily the source rate. Compressed near-full-scale peaks alone do not prove recorded clipping. File extensions are checked first, then actual decoding; this is not a security scanner. Corrupt or unsupported files show an error.
Face visibility: the checklist is manual. Automatic face detection, image restoration, speech recognition and reliable noise detection are outside this MVP. No result guarantees successful face detection or lip sync.
Use the tested examples
The example buttons use the portrait and narration from the settings comparisons. They demonstrate the checker’s measurements; the thresholds have not been validated across a collection of rendered faces.