K Studio LAB

Background Removal Lab

Compare two browser-only portrait models on the same photo. Nothing is uploaded.

Choose a photo to begin.

One photo, all engines

Run the quick browser engines on the same upload. BiRefNet-lite stays available as a manual desktop test because its 114.5 MB load would slow this button. Alpha results use the checkerboard. Rizwan-style uses the same MediaPipe model but white-composites the output, exactly labelled.

Original

Cutout

Model facts

BiRefNet-lite FP16High quality · ~114.5 MB first load · MIT
MODNet quantizedFast portrait mode · ~6.6 MB first load · Apache-2.0
Pas MODNet v1 fine-tuneHair-halo candidate · 25.97 MB · Apache-2.0 v1 weights
MediaPipe Selfie SegmenterInstant preview · 244 KB model · fully on-device · Apache-2.0

AUTO confirms a real WebGPU adapter before choosing BiRefNet-lite; otherwise it chooses MODNet. MediaPipe is the tiny speed baseline. Model files cache after first use.

Real portrait comparison

The same indoor portrait through every available path. Passportphotoedit is shown exactly as delivered: white composite, no alpha output. BiRefNet-lite has no manufactured result - the recorded comparison run failed. AUTO now falls back to MODNet when BiRefNet cannot run.

KStudio four-way background removal comparison grid

Hair and edge detail

Hair and edge detail crops comparing passportphotoedit, MODNet and withoutBG