Dimensions and bytes
Exact file properties. A short side below 256 pixels triggers review; the current per-image upload ceiling is 30 MB.
Private on-device preflight
Check a face swap image before you upload
Choose an image and inspect transparent technical measurements on this device. The checker never uploads the selected file and never converts the metrics into a promise about the generated result.
Select one image
The file stays in this browser tab. It is decoded into a temporary canvas, sampled at no more than 512 pixels on the longest side, and released when you clear or leave the page.
No image selected
JPEG, PNG, WebP, or AVIF supported when the browser can decode it.
No upload: the checker contains no file-transfer request. Normal page telemetry does not receive the file name, image bytes, pixels, or measured values.
Technical measurements
Read the values, then inspect the image
Choose an image to measure it on this device.
No quality score: review flags are conservative prompts, not a model score or a prediction. A clean measurement cannot establish face visibility, identity similarity, consent, pose compatibility, or output realism.
Measurement map
Five signals, each with a narrow meaning
Mean luminance
Average Rec. 709 luma. It can flag a globally dark or bright input but cannot tell whether a face is correctly exposed.
Contrast spread
Population standard deviation of luminance. A low value can indicate a flat input, but a high value is not proof of useful facial detail.
Clipping share
Percentage of sampled pixels near black or white. It identifies compressed tonal range without locating the affected facial features.
Adjacent-edge difference
Mean local luminance change. It responds to blur but also to texture, noise, resizing, and scene content.
Human inspection
Still required for the actual face, pose, expression, hair, hands, glasses, occlusion, permission, and intended audience.
Evidence and limits
See the controlled benchmark behind the analyzer
The published benchmark applies the same formulas to six controlled variants of one owned source. It includes the exact transformations, sample files, JSON, CSV, and SHA-256 hashes. The compressed sample also demonstrates why aggregate metrics must not be presented as a universal pass/fail score.
Checker questions
What the tool does and does not do
Does this upload my image?
No. Decoding, resizing, and measurement happen inside this browser tab.
Does no warning guarantee a strong output?
No. The checker does not inspect the generation model or predict an output.
Does it detect a face?
No. It measures the complete decoded image and does not detect identity, landmarks, pose, expression, or consent.
Why should I still run a small test?
Only a representative generation can show how the selected inputs interact with the current workflow.
Use one representative image before scaling the task
After reviewing the input, open the photo workflow and test the smallest meaningful case first.