mrajaeim/image-pipes
mrajaeim/image-pipes has 133 GitHub stars and 20 forks. GitHot measured +1 stars in 24 hours and +4 in 7 days. Last checked 2026-11-17 02:07 UTC. Last push 2026-11-13.
Measured star growth
- in 24h
- +1
- in 7 days
- +4
- in 15 days
- Collecting history
- in 1 month
- Collecting history
Net star changes over overlapping periods, not daily totals. Missing measurements are left blank. Stars measure attention, not software quality.
About
Image Pipeline Playground
Machine LearningREADME.md
Originalπ¨ Image Pipes
Stop scripting. Start seeing.
Build computer vision pipelines visually β inspect every pixel transformation in real time, then export clean, production-ready Python the moment you're happy with the result.
β‘ Build β’ Preview β’ Experiment β’ Export
Download β’ Demo β’ Features β’ Architecture β’ Docs β’ Quick Start β’ Roadmap
π Experience the Workflow
Visual Editor
|
Dataset AugmentationCreate hundreds of augmentations visually. | Pipeline TemplatesBuild reusable computer vision pipelines. |
π The Problem
If you've built a computer vision pipeline before, this loop probably feels familiar:
tweak a parameter β run the script β save an image β open it β squint β repeat
Each iteration costs you a context switch. Multiply that by every blur kernel, every threshold value, every augmentation you're tuning, and "quick experiment" turns into an afternoon lost to cv2.imshow() windows and throwaway debug scripts.
Image Pipes breaks the loop. Assemble your pipeline as a graph, watch every node's output update live, and walk away with production Python β not a black box you're locked into forever.
ββββββββββββββββ
β Load Image β
ββββββββ¬ββββββββ
βΌ
ββββββββββββββββ
β Resize β
ββββββββ¬ββββββββ
βΌ
ββββββββββββββββ
β Grayscale β
ββββββββ¬ββββββββ
βΌ
ββββββββββββββββ
β GaussianBlur β
ββββββββ¬ββββββββ
βΌ
ββββββββββββββββ
β Threshold β
ββββββββ¬ββββββββ
βΌ
ββββββββββββββββ
β FindContours β
ββββββββ¬ββββββββ
βΌ
Production Python π
β¨ Features
π¨ A Canvas, Not a Console
Assemble pipelines by dragging nodes onto an infinite canvas powered by React Flow. Connect them, configure them through rich auto-generated property panels, and let built-in validation catch mistakes before you ever hit "run."
- Drag & drop nodes with smart, type-aware connections
- Infinite canvas with smooth zoom & pan
- Dynamic property panels generated straight from node metadata
- Real-time validation as you build
β‘ Real-Time Preview, Every Step
Every node executes independently, so you can inspect a transformation the instant it happens β no more mentally simulating what five chained cv2 calls did to your image.
See, live, at every node:
- Intermediate image outputs
- Bounding boxes, segmentation masks, and keypoints
- Execution time per node
- Runtime logs
- Script log lines from custom / reusable Python (
log(...))
π Custom Python & Reusable Scripts
Drop a Custom Python node when a built-in transform isn't enough. Author process(image, seed=0) in the Monaco editor with cv2, numpy, and Image Pipes helpers available in scope:
def process(image, seed=0):
# log(...) appears in the inspector Script log panel
log("input", image, "seed=", seed)
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
return cv2.cvtColor(gray, cv2.COLOR_GRAY2BGR)
Trust before run. Workflows that include custom code are blocked until you review and confirm β trust is session-only and never written into exported workflow JSON.
Save as a reusable node. Promote inline code into a named palette entry under My Scripts. Scripts are versioned on disk (v1, v2, β¦); canvas nodes pin a version so older pipelines keep running the code they were saved with. Editing always creates a new version instead of overwriting history.
Helpers (starting with log(*args)) are documented in-app from the inspector Helpers dialog β more runtime helpers will land there over time. For on-disk layout of reusable scripts and the trust model, see Data locations and Architecture.
π§ A Smart Execution Engine Under the Hood
Pipelines aren't run top-to-bottom blindly. Image Pipes compiles your graph into a Directed Acyclic Graph (DAG), topologically sorts it, and executes only what actually needs to run.
- Incremental execution & partial recomputation
- Run-to-selected-node for fast, targeted debugging
- In-memory caching so large pipelines stay snappy
- Full DAG validation before execution
The result: pipelines with dozens of nodes stay responsive, not sluggish.
π¦ Metadata-Driven Architecture
Adding a new backend node automatically makes it available in the editor β forms, controls, categories, and validation are all generated for free. Zero frontend work required. Implement the node, and the UI builds itself around it.
π§ͺ A Serious Computer Vision Toolbox
Powered by industry-standard libraries, not reinvented wheels.
OpenCV gives you the full classical CV toolkit:
- Color conversions, blur & filtering, edge detection
- Histograms, morphology, geometric transforms
- Contours, thresholding, K-Means, and much more
Albumentations powers first-class data augmentation right inside the visual graph. Chain augmentations like flips, rotations, color jitter, noise, cutout, and elastic transforms alongside your preprocessing nodes β all while correctly propagating:
- Images
- Bounding boxes
- Segmentation masks
- Keypoints
Every run is deterministic and seedable, so your augmented datasets are fully reproducible across teammates and experiments β critical for anyone training a model on top of this pipeline.
π Your Pipeline Is Never Trapped
Design visually, but never get locked in. Export readable, standalone Python built on OpenCV and Albumentations β no proprietary runtime, no generated framework, nothing that only works inside Image Pipes.
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
blur = cv2.GaussianBlur(gray, (5, 5), 0)
edges = cv2.Canny(blur, 50, 150)
That's it. That's the export. Drop it straight into your training pipeline, your ETL job, or a notebook.
π Use Cases
Image Pipes fits naturally into workflows like:
- Dataset preprocessing for ML training
- Building and tuning data augmentation pipelines with Albumentations
- Teaching computer vision concepts visually
- Rapid OpenCV experimentation and prototyping
- Dropping in one-off Custom Python steps (with a trust gate) without leaving the canvas
- Saving reusable, versioned script nodes for your own toolbox
- Research reproducibility (deterministic, versionable pipelines)
- Debugging annotation pipelines (boxes, masks, keypoints)
- General image analysis and exploration
π Architecture
React Flow Editor
β
REST + WebSockets
β
βΌ
DAG Execution Engine
βββββββββββββββΌββββββββββββββ
β β β
βΌ βΌ βΌ
OpenCV Processing Albumentations Custom / user scripts
(classical CV) (augmentation) (process + log helpers)
β β β
βββββββββββββββ΄βββββββ¬βββββββ
β
βΌ
Live Preview & Export
β
βΌ
Standalone Python
π Docs
| Guide | What it covers |
|---|---|
| Architecture & node registration | DAG layers, registering nodes, Custom Python trust gate |
| Data locations | Cache, uploads, outputs, assets, and versioned user_scripts/ on disk (Electron vs backend) |
Also see the desktop notes in desktop/README.md.
π Tech Stack
Frontend
|
Backend
|
Desktop
|
π Quick Start
Download
Grab the latest desktop installer (Windows, macOS, or Linux) from the Releases page β no Python or Node required.
From source
Clone the repository:
git clone https://github.com/mrajaeim/image-pipes.git
cd image-pipes
Install everything:
npm run install:all
Launch the desktop app:
npm run desktop
That's it β the Electron app automatically starts the backend and opens the editor. No separate servers to babysit.
π― Why Developers Love It
| β Visual instead of script-based | β Zero frontend work for custom nodes |
| β Live preview at every step | β Deterministic Albumentations augmentation |
| β Fast, targeted experimentation | β Clean, dependency-free Python export |
| β Metadata-driven architecture | β Open source & self-hosted desktop app |
| β Inline Custom Python when you need it | β Versioned reusable scripts + trust gate |
β
Script log() for in-graph debugging |
β Monaco editor with in-app helper docs |
π£ Roadmap
- π₯ Video pipelines & camera streaming
- π§© ONNX Runtime and PyTorch inference nodes
- β‘ CUDA acceleration
- π¦ Batch processing
- π Plugin SDK (extend beyond built-in + user scripts)
- βοΈ Cloud workspaces
- π€ AI-assisted pipeline generation
- π§° More script helpers beyond
log()(rename / delete / export scripts with workflows)
π€ Contributing
Contributions are welcome β bug fixes, performance improvements, new processing nodes, or entirely new features. If you're planning a large change, please open an issue first so we can discuss the design together.
π License
Distributed under the MIT License. See the LICENSE file for details.