Heavy motion blur, missed focus, very low resolution, and repeated screenshot/compression cycles make hair, fingers, clothing, and product edges harder to separate.
How to get a cleaner browser background cutout.
Automatic segmentation works best when the subject has a clear visual boundary. Source sharpness, contrast, fine hair, transparent materials, motion blur, and compression can all affect the result.
Give the model a clear subject boundary.
A studio wall is not required, but a sharper original image gives the model more useful edge information.
Dark hair on a dark wall, white fabric against a bright sky, glass, smoke, fur, and semi-transparent material are naturally difficult segmentation cases. A different source photo can help more than repeatedly processing the same difficult image.
The current production interface accepts PNG, JPEG, and WebP and checks basic file type and size before browser inference begins.
Why the first run can take longer.
FlytheBG uses IMG.LY on the visitor's device. The browser must obtain the smaller quantized IS-Net model and WebAssembly/runtime software assets before local inference can start.
The current production flow uses IMG.LY's quantized IS-Net model directly. FlytheBG does not automatically download the much larger FP16 model first.
Using the quantized model reduces the automatic model transfer and browser memory pressure while keeping the background-removal workflow on the visitor's device.
Browsers may cache model/runtime software assets. Cache behaviour depends on browser settings, storage availability, private-browsing mode, and later cache eviction.
A returned PNG is not automatically treated as success.
FlytheBG decodes the generated image and checks that it contains visible foreground and a usable transparent area before presenting it as a finished cutout.
A completely or almost completely transparent image is treated as a failed cutout rather than a successful blank checkerboard.
If the generated result has effectively no transparent area, FlytheBG treats the attempt as unusable instead of presenting a false success.
Review hair, ears, shoulders, fingers, product handles, and transparent objects before professional use. Automatic segmentation can still make mistakes.
Process one photo in your browser.
The live remover uses the same smaller browser-only model described above.