pixo 0.3.1

Vision, simplified. Run any computer vision model with one command.

pixo — vision, simplified.

pixo is a runtime layer for computer vision models that lets you run, chain, and manage them as one system on your own machine.

pip install pixo
pixo try

That's it. pixo try auto-picks a model for your hardware, finds a sample image, runs it, and opens a browser report. You go from pip install to a working result in under a minute.

Why pixo?

Running modern computer vision models locally is messy and fragile. pixo makes it safe, consistent, and easy to use multiple models without crashes, conflicts, or glue code.

Problem	Without pixo	With pixo
Sensitive data leaks to cloud	Telemetry and version pings phone home	--airgap blocks every outbound call — provably local
First run takes 10 minutes	Read docs, pick model, find sample, type 4 commands	pixo try — one command, model picked for your hardware
Laptop freezes mid-inference	RAM/CPU maxed out, hard reboot	Resource Guardian caps usage automatically
Job crashes at 80%	Start over from zero	Auto-checkpoint + resume from the last save
Dependency conflicts	Packages break each other	Isolated environments per model
Inconsistent outputs	Every model outputs differently	Standard results.json + COCO / CSV exports
Sharing a result is painful	Zip + instructions + hope it opens	pixo share → single self-contained HTML file

Quick start

pip install pixo

# See it work (60-second hero demo)
pixo try

# Browse models and check your hardware
pixo list
pixo doctor

# Run any model on your own file
pixo run yolov8 --input photo.jpg
pixo run yolov8 --input video.mp4

# Speed up with ONNX (~40% faster on CPU)
pixo optimize yolov8

Supported models

Every model carries a privacy badge — all nine bundled models run fully offline after weights are downloaded once.

Model	Task	Default size
yolov8 / yolov11 / yolov12	Object detection	5–7 MB
rtdetr	Transformer detection	64 MB
grounding_dino	Text-prompted detection	341 MB
florence2	Vision-language (caption, detect, OCR)	460 MB
depth_anything_v2	Depth estimation	99 MB
sam2	Segmentation	898 MB
samurai	Video tracking + segmentation	898 MB

pixo run grounding_dino --input photo.jpg --prompt "red backpack, yellow hat"
pixo run florence2 --input photo.jpg --task caption
pixo run depth_anything_v2 --input photo.jpg

Key features

Local-first by design

pixo run yolov8 --input photo.jpg --airgap

Blocks all outbound network calls for the duration of the run — proves your data never left the machine. Combined with privacy badges on every model card, pixo is the only CV runtime built around "no bytes leave your laptop."

See where models disagree

pixo compare yolov8 yolov11 yolov12 --input photo.jpg

Runs multiple detection models on the same image and shows only where they disagree. Agreement / partial / unique detections visualized side-by-side in a shareable HTML report.

Shareable reports, no server

pixo share

Exports a run as a single self-contained .html file with images embedded. Attach it to a tweet, Slack, or email — anyone can open it in any browser.

Safe on any laptop

pixo run yolov8 --input video.mp4 --low-memory --background

Resource Guardian caps RAM, CPU, and GPU usage. --background drops priority so your laptop stays responsive while pixo runs.

Never lose progress

pixo run yolov8 --input long_video.mp4
# [Ctrl+C pauses, saves checkpoint]
pixo resume

Auto-saves every N frames. Ctrl+C pauses gracefully instead of killing the process.

Free cloud GPUs

pixo setup-cloud --kaggle
pixo run sam2 --input photo.jpg --backend kaggle

Route heavy models to free Kaggle or Colab GPU. 30 hours/week free on Kaggle.

Chain models

pixo pipe "grounding_dino -> sam2" --input photo.jpg --prompt "person"

Browser UIs (optional)

pip install pixo[demo]
pixo serve yolov8          # Gradio UI for one model

pip install pixo[web]
pixo ui                    # Full local dashboard

All commands

pixo try                   # Zero-setup hero demo
pixo list                  # List models with privacy badges
pixo info <model>          # Detailed model info
pixo pull <model>          # Pre-download a model
pixo run <model> -i <f>    # Run inference
pixo compare <m1> <m2> ... # Cross-model disagreement browser
pixo share [job_id]        # Export self-contained HTML report
pixo serve <model>         # Gradio browser UI
pixo ui                    # Full web dashboard
pixo pipe "m1 -> m2"       # Chain models
pixo doctor                # Check hardware
pixo optimize <model>      # ONNX conversion (~40% faster CPU)
pixo history               # Show past jobs
pixo resume [job_id]       # Resume a paused job
pixo view <job_id>         # Open a job's output folder
pixo setup-cloud           # Connect Kaggle / Colab
pixo cloud-status          # Cloud backend status
pixo rm <model>            # Remove a downloaded model
pixo upgrade               # Update pixo
pixo guide                 # In-terminal usage guide

Python SDK

import pixo

result = pixo.run("yolov8", input="photo.jpg")
print(result.objects, result.classes, result.time_seconds)

result = pixo.run("grounding_dino", input="photo.jpg", prompt="red car")
result = pixo.run("sam2", input="photo.jpg", backend="kaggle")

hw = pixo.doctor()
print(hw["ram_total_gb"], hw["has_gpu"])

Adding a model

pixo uses a plugin system. Each model is two files:

pixo/models/cards/your_model/
  modelcard.yaml   # Metadata, dependencies, hardware, privacy
  run.py           # Two functions: setup() and run()

Optional installs

pixo keeps the base install minimal. Pull only what you need:

pip install pixo[yolo]     # YOLO family (Ultralytics + OpenCV)
pip install pixo[onnx]     # ONNX Runtime for faster CPU
pip install pixo[vision]   # Grounding DINO / Florence-2 / SAM2
pip install pixo[cloud]    # Kaggle backend
pip install pixo[demo]     # Gradio for pixo serve
pip install pixo[web]      # FastAPI + uvicorn for pixo ui
pip install pixo[all]      # Everything

Development

git clone https://github.com/Janinduu/pixo.git
cd pixo
pip install -e .

Documentation

Detailed feature guides live in docs/:

• Features summary — one-page reference
• Features explained — full guide with why / what / how

License

MIT