Describe
One typed Style
Box, stroke, fill, line, label, effects and trail live in a single Pydantic model. Every choice is inspectable, portable, and saves as plain YAML.
Object-detection styling for Python
Boxes that
look authored.
visionstyle turns raw model output into a deliberate visual language. One typed
Style describes boxes, labels, fills, glow, glass, grain and tracking trails —
tuned live in the Studio, saved as YAML, rendered with one call.
- 12built-in presets
- 1typed
Style - 3.10+Python
- MITlicence
live preview · night.jpg · yolo11n.pt
vs.annotate(frame, dets, style="cinematic")
drag to compare
01 / The idea
Technical sight, turned into a visual language.
Most pipelines stop at a rectangle. visionstyle starts there — then gives you a complete, composable system for shape, label, colour, motion, texture and tracking, all described by one object.
Tune
By eye, in the Studio
Your image, your model, live controls. The preview is rendered by the Python package itself, so what you see is exactly what annotate() produces.
Track
Let it move
Keep one Annotator per stream and trails, marching dashes, hue cycles and pulses carry across frames — on video or a live webcam.
Render
Any resolution
Sizes are reference pixels at ~1080p and scale automatically, so one style looks identical on a webcam frame and a 4K deliverable.
02 / Presets
Twelve starting points. Every one is YAML.
Pick a mood, then make it yours. Each preset is a plain file you can copy, edit and load by name — vs.Style.preset("hud").
Amber / teal rounded frames, pill labels, soft glow, filmic grade
style = vs.Style.preset("cinematic")
03 / Configure
Every parameter. One Style.
Start from a preset, a YAML path, or nothing at all. Every option is a typed attribute, so your editor autocompletes it and Pydantic validates it.
# detections: boxes in pixels, everything else optional
dets = vs.Detections(
xyxy=[[590, 650, 720, 1040], [1060, 660, 1520, 1000]],
class_name=["person", "car"],
confidence=[0.93, 0.88],
track_id=[14, 31],
)
# or: vs.Detections.from_ultralytics(model.predict(frame)[0])
style = vs.Style.preset("cinematic")
style.label.components = ["track_id", "text"]
style.trail.enabled = True
annotator = vs.Annotator(style) # one per stream
out = annotator.annotate(frame, dets)
style.save_preset("my-look") # ~/.visionstyle/presets/my-look.yaml# src/visionstyle/presets/cinematic.yaml (excerpt)
name: cinematic
palette: {colors: cinematic, by: class}
box: {shape: rounded, corner_radius: 10}
stroke: {thickness: 2}
fill: {enabled: true, opacity: 0.1, mode: gradient}
label:
components: [text, confidence]
background: pill
font_weight: 600
effects:
glow: {enabled: true, radius: 10, intensity: 0.55}
vignette: {enabled: true, strength: 0.5, radius: 0.9}
grain: {enabled: true, amount: 0.04}
trail: {enabled: true, length: 36, line: ribbon, glow: true}- Box
- rectangleroundedcornersreticlenonedouble linecenter mark
- Stroke
- thicknessopacitypaletteconfidence rampany hex / rgb / name
- Fill
- solidgradient · 5 directionshatchopacity
- Line
- soliddasheddottedsegmentsperimeter gradientmarchhue_cyclepulse
- Label
- textconfidencetrack_idcustom template9 anchors × in / outvertical tagsInter · JetBrains Mono · your .ttfsolidpillglassunderline
- Effects
- glowshadowfrosted glassspotlightvignettefilm graincolour grade
- Tracking
- trail lengthfeet / center / topsoliddotteddashedribbonfade & tapersmoothingpoints
- Global
- paletteper-class overridesconfidence thresholdscale = autofps
04 / The Studio
The preview is the hero,
not a widget.
Design a style visually, on your own image and model, and save it as a preset the package loads by name. Nothing is faked: the canvas is rendered by the same Python engine you ship.
- Sources
Upload an image or use the bundled samples; bring a
.pt/.onnxmodel or start withyolo11n.pt. - Controls
Every section of the Style is a live control. Toggle objects, isolate one, play line animations, preview trails on a still.
- Export
YAML with only changed values or everything, a clean Python snippet, or Save as preset straight into your presets directory.
pip install "visionstyle[studio,yolo]"
visionstyle studio # → http://127.0.0.1:842005 / Install
A few lines
from beautiful.
Core depends on NumPy, OpenCV-headless, Pillow, Pydantic and PyYAML. Add extras for Ultralytics models or the Studio.
-
1Install the package
pip install visionstyleThe latest release from PyPI. Python 3.10 – 3.13.
pip install "visionstyle[yolo]" # + ultralytics for demos / CLI models pip install "visionstyle[studio]" # + fastapi / uvicorn for the Studio pip install "visionstyle[all]"pip install git+https://github.com/baselhusam/visionstyle.gitThe unreleased development version from
main.visionstyle render photo.jpg -s neon --model yolo11n.pt -o out.jpg visionstyle render clip.mp4 -s tracking --model yolo11n.pt --track -o out.mp4 visionstyle render 0 -s hud --model yolo11n.pt --track -o webcam.mp4 visionstyle gallery photo.jpg --model yolo11n.pt -o gallery.pngRender stills, videos or a webcam index;
galleryrenders every preset at once. -
2Style your first frame
import cv2 import visionstyle as vs from ultralytics import YOLO frame = cv2.imread("street.jpg") result = YOLO("yolo11n.pt")(frame)[0] dets = vs.Detections.from_ultralytics(result) cv2.imwrite("out.jpg", vs.annotate(frame, dets, style="cinematic"))Needs the
[yolo]extra. No model handy?visionstyle render sample -s neon -o out.jpgrenders a bundled photo with its shipped detections.