⚡ VisiMac Studio

High-Performance Computer Vision & Dataset Preparation Engine

VisiMac Studio is an all-in-one workstation designed for computer vision engineers, AI researchers, and data scientists. It bridges the gap between raw data collection and production model deployment by integrating multi-unit image batch processing, grid tile splitting, synthetic data augmentation, interactive dataset annotation, AutoML hyperparameter optimization, YOLO training/metrics tracking, and model export.

🛠️ Installation & Launch

Install the pre-compiled binary package directly using pip:

pip install visimac

Launch the workstation GUI directly from your terminal:

visimac

🚀 Workstation Guide

VisiMac Studio is designed as a sequential pipeline. The tabs at the top of the interface follow the natural progression of computer vision model development, from raw image processing to final model deployment.

1. Image Resizer + Compressor

Prepare raw, high-resolution source images for annotation or direct model ingestion without sacrificing aspect ratios or ballooning storage sizes.

• Dimensional Scaling: Resize entire directories using exact pixel bounds or physical print units (cm, mm, in, pt) mapped to target DPI settings. Include subfolders recursively to process large directory trees.
• Exact Target Size Compression: Check "Enable Target Size Constraint" and define a maximum file size in kilobytes (e.g., 500 KB). The internal binary search algorithm calculates the mathematically optimal JPEG/WEBP quality factor for each individual image to hit the target file size without losing excess visual fidelity.
• Format & Rotation Batching: Convert image pools dynamically to JPEG, PNG, or WEBP while applying fixed orientation corrections (90°, 180°, 270°).
• Live Queue Analytics: Select any file in the processing queue to view a real-time estimation of the output resolution, DPI, and pre-compression source size.

2. Image Splitter

Decompose massive resolutions (such as satellite imagery, medical scans, or industrial panoramas) into manageable grids to maintain native detail during YOLO training.

• Dynamic Grid Layout: Specify the total number of target tiles. The engine calculates the optimal row-by-column matrix (e.g., entering 4 generates an auto-balanced 2 Rows x 2 Cols split).
• Structured Outputs: Toggle subfolder sorting to save each extracted tile into organized directories (e.g., part_1/, part_2/), preventing flat-folder clutter and aiding in parallel dataset construction.

3. Synthetic Data Enricher

Multiply your dataset variance and correct perspective distortions before object annotation.

• Interactive Transformation Canvas: Select an image and use the active mode dropdown to apply destructive spatial operations. Select Crop to isolate objects, Warp to map four corner points and flatten perspective distortion, or Rotate (fixed 90-degree steps or custom precision angles) to correct misaligned sensor data.
• Augmentation Ranges: Define strict Min/Max tolerances for visual variance, including:
  • Brightness Additions (-100 to +100 intensity offset)
  • Contrast Scaling (0.1x to 3.0x multiplier)
  • Rotation Degrees (slight rotational variance)
  • Lighting/Gamma Curve Adjustments
  • Perspective Warp Skew Variables
• Live Visual Simulation: Click "Simulate Augmentation Canvas" to immediately preview how the engine (and subsequently YOLO) will distort your source data.
• Batch Variant Generation: Set "Variants per Source" to instantly synthesize and export multiple unique, randomized images from your defined ranges.

4. Image Annotator

A dedicated, distraction-free labeling environment for defining bounding boxes and segmentation masks.

• Multi-Shape Draw Tools: Trace targets using bounding Rectangles or segmentation Polygons, Circles, and Pentagons.
• AI Auto-Labeling Assist: Load pre-trained or previously generated YOLO model weights (.pt). Adjust the confidence threshold and click "Auto-Label Frame" to automatically draw bounds and masks over detected objects.
• Sequence Frame-Copying: Click "Copy Previous Frame Shapes" when labeling rapid video frames or stationary objects to instantly duplicate the prior image's exact coordinates.
• Semi-Transparent Overlays: Shapes are automatically filled with distinct, semi-transparent colors mapped to specific class IDs for clear overlap visibility. Selected shapes boost to 100% opacity for precise coordinate adjustments.
• Dataset Health Inspector: Run a comprehensive QA scan against the loaded directory. The inspector flags unannotated images, orphaned .txt files without matching images, and duplicate/overlapping annotations where the Intersection-over-Union (IoU) exceeds 90%.
• Export & Split Dataset: Select your target split ratio (e.g., 70/20/10 for Train/Val/Test). The engine automatically shuffles the data, copies images and labels to the appropriate subdirectories, standardizes YOLO coordinates, and generates a valid data.yaml manifest.

5. YOLO Configuration & AutoML

Define your model architecture and discover mathematically optimal training settings.

• Model & Dataset Mapping: Select from standard YOLO backbones (e.g., yolov8n.pt, yolov8x-seg.pt). Map the output directory and select the data.yaml generated in the previous step.
• Manual Hyperparameters: Configure Epochs, Batch Size, Input Image Size, Optimizer (AdamW, SGD, Adam), Data Loader Worker Threads, Initial/Final Learning Rates, and Mosaic thresholds.
• AutoML Genetic Search: Instead of guessing hyperparameter configurations, click "Optimize Hyperparameters". This triggers a background routine that runs multiple abbreviated training lifecycles, cross-mutating variables to determine the mathematically optimal setup for your specific dataset, automatically applying the winning values back to the GUI.

6. Training Analytics & Live Logs

Monitor model convergence in real time without executing terminal commands.

• Console Stream: View real-time standard output and error logging directly inside the workstation.
• Live Loss Curves: The live plot graphs Train Box Loss, Validation Box Loss, and Train Class Loss dynamically as epochs complete.
• Diagnostic Rendering: Following training, visually inspect model performance via dedicated tabs for the generated F1-Score Curve, Precision-Recall Curve, Confusion Matrix, and overlaid Validation Batch Predictions.
• Safe Halting: Click "Stop Training" to gracefully conclude the training run at the end of the active epoch without corrupting the saved weights.

7. Inference & Model Export

Verify the newly trained model against isolated test data and compile it for production hardware.

• Batch Inference: Load your best .pt file and select a target folder of raw images. Dial in confidence thresholds and image input sizes.
• Prediction Visualizer: Navigate sequentially through the inference queue to review object detections, boundary precision, and classification confidence values drawn directly on the frames.
• Production Deployment Export: Select a target save directory and convert PyTorch weights (.pt) into optimized deployment formats:
  • ONNX (Cross-platform compatibility)
  • engine (NVIDIA TensorRT for maximum GPU acceleration)
  • openvino (Intel CPU/iGPU optimization)
  • torchscript (C++ production environments)
  • coreml (Apple Silicon iOS/macOS Neural Engine)

💻 Cross-Platform Support