cvPal
cvPal

Core Features

Synthetic Generation

Generate synthetic datasets with automatic object detection and labeling using AI models.

Overview

cvPal's synthetic generation feature combines state-of-the-art image generation models (DALL-E 3, Stable Diffusion) with automatic object detection (OWL-ViT) to create fully annotated datasets. Simply provide a text prompt and labels, and cvPal will generate images with precise bounding box annotations.

How It Works

1

Generate Image

AI model creates image from text prompt

2

Detect Objects

OWL-ViT identifies objects in image

3

Create Labels

Generate bounding box annotations

4

Save Dataset

Export in YOLO or COCO format

Basic Usage

Get started with synthetic dataset generation using the DetectionDataset class:

Simple Example

python
from cvpal.generate import DetectionDataset


# Initialize with Stable Diffusion (default)
dataset = DetectionDataset(model="stable-diffusion")


# Generate a dataset
dataset.generate(
    prompt="a cat sitting on a chair",
    num_images=5,
    labels=["cat", "chair"],
    output_type="yolo"
)

📝 Parameters

  • prompt - Text description for image generation
  • num_images - Number of images to generate
  • labels - List of object classes to detect
  • output_type - "yolo" or "coco" format
  • height/width - Image dimensions (default: 512x512)

🎯 Output

  • • Generated images in images/ folder
  • • Corresponding labels in labels/ folder
  • • Dataset configuration in data.yaml
  • • Optional COCO annotations in annotations.json

Model Selection

Choose between different image generation models based on your needs:

Stable Diffusion (Recommended)

Free, local processing with high customization options. Best for most use cases.

python
# Stable Diffusion - No API key required
dataset = DetectionDataset(model="stable-diffusion")


# Customize generation parameters
dataset.generate(
    prompt="a dog playing in the park",
    num_images=10,
    labels=["dog", "park", "tree"],
    height=768,
    width=768,
    seed=42,
    output_type="yolo"
)

DALL-E 3

Premium quality images with excellent prompt understanding. Requires OpenAI API key.

python
# DALL-E 3 - Requires OpenAI API key
dataset = DetectionDataset(
    model="dalle",
    openai_api_key="your-openai-api-key"
)


# Generate high-quality images
dataset.generate(
    prompt="a professional photo of a cat in a business suit",
    num_images=3,
    labels=["cat", "suit"],
    output_type="coco"
)

Advanced Features

Dataset Management

Manage and extend your generated datasets with built-in utilities:

Add Labels to Existing Dataset

python
# Add new labels to existing images
dataset.add_labels(labels=["person", "car"])


# This will:
# - Run detection on all existing images
# - Add new annotations to label files
# - Update data.yaml with new classes

Quality Control

python
# Check for images with no detections
dataset.isnull()


# Remove images with no detections
dataset.dropna()


# Visualize samples
dataset.show_samples(num_samples=5, annotation_type="yolo")

Batch Processing

Generate multiple datasets efficiently with parallel processing:

python
# Generate multiple datasets
prompts = [
    "a cat sitting on a chair",
    "a dog playing with a ball",
    "a bird flying over trees"
]


for i, prompt in enumerate(prompts):
    dataset = DetectionDataset(model="stable-diffusion")
    dataset.generate(
        prompt=prompt,
        num_images=5,
        labels=["animal", "object"],
        output_type="yolo",
        overwrite=True
    )
    print(f"Generated dataset {i+1}/3")

Output Formats

Choose the output format that best fits your training pipeline:

YOLO Format

Each image gets a corresponding .txt file with normalized coordinates.

text
# image_001.txt
0 0.5 0.3 0.2 0.4  # cat: class_id x_center y_center width height
1 0.7 0.6 0.15 0.3  # chair: class_id x_center y_center width height

Best for: YOLOv5, YOLOv8, custom detection models

COCO Format

Single JSON file with comprehensive metadata and annotations.

json
{
  "images": [{"id": 1, "file_name": "image_001.jpg", "width": 512, "height": 512}],
  "annotations": [{"id": 1, "image_id": 1, "category_id": 1, "bbox": [100, 50, 200, 150], "area": 30000}],
  "categories": [{"id": 1, "name": "cat"}]
}

Best for: Detectron2, MMDetection, COCO evaluation

Best Practices

✅ Effective Prompts

  • • Be specific about object types and positions
  • • Include environmental context
  • • Use descriptive adjectives
  • • Mention lighting and style preferences
  • • Avoid ambiguous descriptions

Good: "a black cat sitting on a wooden chair in a living room"
Bad: "cat and chair"

🎯 Label Strategy

  • • Use consistent naming conventions
  • • Include all objects you want to detect
  • • Consider hierarchical labels
  • • Test detection threshold (0.1-0.3)
  • • Validate annotations manually

Example: ["person", "car", "traffic_light", "road"]

Performance Tips

🚀 Speed

  • • Use GPU acceleration
  • • Reduce inference steps
  • • Lower detection threshold
  • • Use parallel processing
  • • Consider smaller image sizes

💾 Memory

  • • Process images in batches
  • • Use torch.float16 precision
  • • Clear GPU cache regularly
  • • Monitor memory usage
  • • Use CPU fallback if needed

🎨 Quality

  • • Use higher resolution images
  • • Increase inference steps
  • • Fine-tune detection threshold
  • • Use diverse prompts
  • • Validate with manual inspection

Troubleshooting

No Objects Detected

If OWL-ViT doesn't detect objects in generated images:

  • • Lower the detection threshold (try 0.05-0.1)
  • • Use more specific labels in your prompt
  • • Ensure labels match objects in the image
  • • Try different image generation models

Poor Image Quality

For better image generation results:

  • • Increase inference steps (50-100)
  • • Use higher resolution (768x768 or 1024x1024)
  • • Improve prompt specificity
  • • Consider using DALL-E 3 for premium quality

Memory Issues

If you encounter GPU memory errors:

  • • Reduce batch size or image count
  • • Use smaller image dimensions
  • • Enable CPU fallback
  • • Clear GPU cache between generations
Supported ModelsDataset Merging