Background removal

BrandKit removes backgrounds with rembg running an ONNX U²-Net model locally, through onnxruntime. Nothing is sent to an external service.

Turning it on

Tick Remove background in the preprocessing panel. It is off by default, and the control is hidden entirely if rembg failed to import at startup — check the logs for:

Background removal (rembg) not available. Install with: pip install rembg

Choosing a model

The method selector maps to a different ONNX model:

Method	Model	Use it for
auto (default)	u2net	general-purpose — logos, products, most things
object	u2net	same as auto, named explicitly
person	u2net_human_seg	photographs of people; much better at hair and limbs
anime	u2net_anime	illustrated and cel-shaded artwork

There is no automatic detection behind auto — it simply uses the general model. If your subject is a person, pick person explicitly; the difference is large.

The first run downloads ~180 MB

On first use of each model, rembg fetches the .onnx file into ~/.u2net/. That request needs outbound internet access and will make the first generation take a minute or more. Subsequent runs are fast. In Docker, mount $HOME/.u2net:/root/.u2net so the download survives image rebuilds — see Running with Docker.

What you get

The output is always RGBA with a real alpha channel. What happens next depends on the background colour setting:

background_color	Result
transparent (default when removal is on)	alpha is preserved — PNG and WebP keep it, JPG flattens to white
a hex value like #0f172a	the cutout is composited onto that solid colour
the detected prominent colour	the smart fill path, driven by image analysis

Hex parsing accepts #abc, #aabbcc, and the same without the leading #. Anything it cannot parse falls back to white rather than failing the request.

JPG and transparency do not mix

If you select jpg as an output type with a transparent result, the alpha is flattened. Select PNG or WebP if you need the cutout to stay cut out.

Cleaning up the edges

Model output is rarely pixel-perfect. Three controls help:

• Edge smoothing (edge_smooth, radius default 2) — feathers the alpha channel so the cutout does not have a jagged staircase against a contrasting background.
• Auto-crop (auto_crop) — after removal the subject is usually surrounded by fully transparent pixels. Auto-crop trims to the bounding box, which makes the subject fill each format instead of floating in the middle.
• Drop shadow (shadow_effect) — only does anything useful once there is an alpha channel, so it pairs naturally with removal.

A good default recipe for a logo on a white JPEG:

remove_background = true
background_removal_method = auto
background_color = transparent
edge_smooth = true, smooth_radius = 2
auto_crop = true, crop_padding = 16
output_formats = png, webp

When it goes wrong

Symptom	Cause	Fix
Nothing happens, background still there	rembg not importable	check startup logs; pip install rembg onnxruntime
First request hangs, then works	model download	expected; pre-seed ~/.u2net/
Parts of the logo were eaten	U²-Net treated flat brand colour as background	try object, or skip removal and use the smart white-fill instead
Halo of white pixels around the subject	source was a JPEG with compression artefacts against white	enable edge smoothing; start from a PNG if you have one
Process killed mid-generation	ONNX Runtime memory spike on a large image	shrink the source, or raise the container memory limit — see Performance

Running without it

rembg and onnxruntime are the two heaviest dependencies in the project — together they account for most of the ~2 GB Docker image. If you never need background removal, remove both lines from requirements.txt before installing. The app detects their absence at import time and degrades gracefully; every other feature is unaffected.