project: projects/useknockout-python type: readme

🥊 useknockout

State-of-the-art background removal API — Python SDK.

Website · API repo · Node SDK · React SDK · CLI

Background removal, 40× cheaper than remove.bg. MIT licensed. Self-hostable.

What it is

useknockout is the official Python SDK for the useknockout background-removal API — a BiRefNet-powered image service that runs on Modal. Beyond plain cutouts, the API exposes a full suite of subject-aware image operations: background replacement, drop shadows, stickers and outlines, e-commerce / LinkedIn presets, super-resolution upscaling, face restoration, colorization, silhouettes, and inpainting.

The SDK is a thin, fully-typed wrapper over httpx. It ships both a synchronous client (Knockout) and an asynchronous one (AsyncKnockout) with identical method signatures. Every image-returning method hands you raw bytes, so you can pipe results straight to a file, BytesIO, PIL, numpy, S3, or anywhere else bytes go. There are no heavy dependencies — no Pillow, no numpy — just httpx.

Install

pip install useknockout

Requires Python 3.9+. The only runtime dependency is httpx>=0.27.0.

Quickstart

from useknockout import Knockout

client = Knockout()  # uses public beta token by default
png = client.remove("photo.jpg")

with open("out.png", "wb") as f:
    f.write(png)

That's it. Free during the public beta.

Async

AsyncKnockout mirrors the sync API one-for-one. Use it as an async context manager so the underlying httpx.AsyncClient is closed cleanly:

import asyncio
from useknockout import AsyncKnockout

async def main():
    async with AsyncKnockout() as client:
        png = await client.remove("photo.jpg")
        open("out.png", "wb").write(png)

asyncio.run(main())

The sync client is also a context manager (with Knockout() as c: ...) and exposes close() / aclose() if you prefer to manage the lifecycle manually.

Authentication

The client resolves its token by falling through three sources, in order:

1. The token= constructor argument
2. The KNOCKOUT_TOKEN environment variable
3. The public beta token (free, rate-limited — baked into the SDK)

client = Knockout(token="kno_...")          # explicit
client = Knockout()                          # env var or public beta token

The token is sent as a Bearer header on every request. Get your own token by emailing troy@useknockout.com (paid tier launches at $0.005/image — 40× cheaper than remove.bg).

Input types

Every method that takes an image accepts a FileInput, which is any of:

• str / pathlib.Path — a filesystem path the SDK reads for you
• bytes / bytearray — raw image bytes you already have in memory

client.remove("photo.jpg")                   # path
client.remove(Path("photo.jpg"))             # Path
client.remove(open("photo.jpg", "rb").read())# bytes

All endpoints

from useknockout import Knockout

c = Knockout()

# --- Core background removal ---
c.remove("photo.jpg")                                  # returns PNG/WebP bytes
c.remove_url("https://example.com/photo.jpg")          # fetch + remove remote image
c.remove_batch(["a.jpg", "b.jpg", "c.jpg"])            # up to 10 files -> JSON
c.remove_batch_url(["https://...", "https://..."])     # up to 10 URLs -> JSON

# --- Replace background ---
c.replace_background("photo.jpg", bg_color="#000000")
c.replace_background("photo.jpg", bg_url="https://example.com/bg.jpg")

# --- Mask / crop ---
c.mask("photo.jpg")                                    # alpha mask only (grayscale PNG)
c.smart_crop("photo.jpg", padding=24)                  # crop to subject bbox + padding

# --- Effects ---
c.shadow("photo.jpg", shadow_blur=14, shadow_opacity=0.45)   # drop shadow
c.sticker("photo.jpg", stroke_width=20)                # WhatsApp-style thick outline
c.outline("photo.jpg", outline_color="#000000", outline_width=4)  # thin outline
c.silhouette("photo.jpg", subject_color="#7C3AED")     # two-tone silhouette portrait

# --- Presets ---
c.studio_shot("photo.jpg", aspect="1:1", shadow=True)  # e-commerce
c.studio_shot("photo.jpg", transparent=True)           # transparent bg (PNG)
c.studio_shot("photo.jpg", enhance=True)               # brightness + saturation lift
c.headshot("photo.jpg", bg_color="#FFFFFF")            # LinkedIn 4:5 portrait
c.headshot("photo.jpg", bg_blur=True, blur_radius=24)  # blurred original bg

# --- Enhancement ---
c.upscale("photo.jpg", scale=4, model="swin2sr")       # 2x/4x super-resolution
c.upscale("photo.jpg", model="realesrgan", face_enhance=True)  # legacy + GFPGAN
c.face_restore("portrait.jpg")                         # GFPGAN v1.4 face restoration
c.colorize("bw.jpg")                                   # DDColor colorization
c.inpaint("photo.jpg", bbox=(40, 40, 200, 150))        # LaMa inpainting

# --- Marketing ---
c.compare("photo.jpg")                                 # before/after side-by-side

# --- UX helpers ---
c.preview("photo.jpg", max_dim=512)                    # fast low-res preview (~80ms warm)
c.estimate("remove", width=1024, height=1024)          # predict latency + cost (no processing)

# --- Telemetry ---
c.health()                                             # service status + model info
c.stats()                                              # public usage counter

Most methods return raw image bytes. The batch endpoints (remove_batch, remove_batch_url) and the telemetry / estimate endpoints (health, stats, estimate) return parsed JSON (dict). Pipe image bytes to open(path, "wb").write(...), BytesIO, PIL.Image.open(BytesIO(...)), numpy, S3 — anywhere bytes go.

Endpoint reference

Enhancement endpoints in detail

upscale — 2x or 4x super-resolution. model="swin2sr" (default) is sharper on real photos; model="realesrgan" is the legacy backend, better on anime / illustrations. face_enhance=True routes through GFPGAN (Real-ESRGAN backend only). scale must be 2 or 4; model must be "swin2sr" or "realesrgan" — both are validated client-side and raise ValueError otherwise.

face_restore — GFPGAN v1.4 portrait restoration for blurry / damaged / low-res faces. By default the background is preserved as-is; bg_enhance=True also upscales the background 2x via Real-ESRGAN. only_center_face=True restores just the most prominent face (faster).

colorize — DDColor (Apache-2.0) colorization via a ConvNeXt-Large backbone (single feed-forward, ~500ms warm). Works on B&W or color input — color images are converted to grayscale internally first, making round-trip recoloring easy.

inpaint — LaMa-based inpainting with three auto-detected modes:

c.inpaint("photo.jpg", mask="mask.png")             # user-supplied mask
c.inpaint("photo.jpg", bbox=(x, y, w, h))           # rectangular region
c.inpaint("photo.jpg")                              # auto-subject (BiRefNet, inverted)

dilation (0–32, default 8) expands the mask before LaMa runs to reduce ghost outlines.

Errors

All errors inherit from KnockoutError. The SDK maps HTTP status codes to typed subclasses, and wraps transport-level failures (httpx.RequestError) as a base KnockoutError with code="unknown".

from useknockout import Knockout
from useknockout.errors import (
    AuthError,           # 401 / 403
    BadRequestError,     # 400
    PayloadTooLargeError,# 413 (image too large)
    RateLimitError,      # 429
    ServerError,         # 5xx
    KnockoutError,       # base class / network errors
)

try:
    png = client.remove("photo.jpg")
except AuthError:
    ...  # bad token
except RateLimitError:
    ...  # back off
except KnockoutError as e:
    print(f"{e.code} ({e.status}): {e.message}")

Each KnockoutError carries three attributes: status (HTTP code, or None for transport errors), code (a short machine-readable category like "auth" or "rate_limit"), and message (human-readable detail, taken from the response body when available).

Configuration

Self-host

The API is open source and runs on Modal. Deploy your own copy:

git clone https://github.com/useknockout/api
cd api
modal deploy main.py

Then point the SDK at your deployment:

client = Knockout(token="your-token", base_url="https://your-deploy.modal.run")

Project structure

useknockout-python/
├── pyproject.toml              # hatchling build, deps, ruff config
├── README.md
├── LICENSE                     # MIT
├── dist/                       # built wheel + sdist artifacts
└── src/
    └── useknockout/
        ├── __init__.py         # public exports (Knockout, AsyncKnockout, errors)
        ├── _version.py         # __version__
        ├── client.py           # synchronous Knockout client
        ├── async_client.py     # asynchronous AsyncKnockout client
        ├── errors.py           # KnockoutError hierarchy + raise_for_status
        └── _helpers.py         # multipart/form helpers, token resolution, constants

The package uses a src/ layout and is built with Hatchling. Both clients share _helpers.py (multipart construction, form serialization, token resolution) and errors.py (status-to-exception mapping), keeping the sync and async surfaces in lockstep.

Development

pip install -e ".[dev]"     # installs pytest, pytest-asyncio, ruff, mypy
ruff check .                # lint (line length 100; E, F, I, W, UP rules)
mypy src                    # type-check
pytest                      # run tests

Notes

• No heavy deps. httpx is the only runtime dependency. Image manipulation happens server-side; the SDK just moves bytes.
• Bytes in, bytes out. Image-returning methods give you raw bytes — no implicit Pillow / numpy conversion, so you stay in control of the output pipeline.
• Sync + async parity. Knockout and AsyncKnockout have identical method signatures; switch by adding/removing await.
• API version gating. Some methods require a minimum API version: colorize needs API ≥ v0.7.0, silhouette ≥ v0.7.1, inpaint ≥ v0.8.0. The default hosted endpoint is kept current.
• format quirk. remove and remove_batch pass format as a query parameter (the API reads it there), while form-based endpoints send it in the multipart body. The SDK handles this for you.

License

MIT. Use it however you want.