Generates synthetic object manipulation tasks where specific objects must be removed from a scene based on various criteria such as color, shape, position, or uniqueness. All other objects remain unchanged in their original positions.

Each sample pairs a task (first frame + prompt describing what needs to happen) with its ground truth solution (final frame showing the result + video demonstrating how to achieve it). This structure enables both model evaluation and training.

# Clone the repository
git clone https://github.com/VBVR-DataFactory/O-43_object_subtraction_data-generator.git
cd O-43_object_subtraction_data-generator

# Install dependencies
pip install -r requirements.txt

# Generate 100 samples
python examples/generate.py --num-samples 100

# Generate with specific seed
python examples/generate.py --num-samples 100 --seed 42

# Generate without videos
python examples/generate.py --num-samples 100 --no-videos

# Custom output directory
python examples/generate.py --num-samples 100 --output data/my_output

Remove the leftmost objects from the scene. Keep all other objects unchanged.

Initial Frame Multiple colored objects in scene	Animation Specified objects disappearing	Final Frame Remaining objects unchanged

Remove specific objects from a scene based on selection criteria (color, shape, position, or uniqueness) while keeping all other objects stationary and unchanged.

• Object Count: 5-8 geometric objects per scene
• Object Types: Cubes, spheres, pyramids, cones
• Colors: Various colors (red, green, blue, yellow, orange, purple)
• Four Task Types:
  • Type 1: Remove by attribute (all objects of a specific color or shape)
  • Type 2: Remove specific objects (one or more named objects like "red cube")
  • Type 3: Remove by position (leftmost, rightmost, N leftmost, etc.)
  • Type 4: Remove the different one (object that looks different from others)

• Attribute-based selection: Tests understanding of color and shape properties
• Specific object identification: Requires matching exact object descriptions
• Spatial reasoning: Identifies objects by relative position
• Outlier detection: Finds objects that differ from the group
• Selective removal: Only specified objects disappear
• Position preservation: Remaining objects stay in exact original locations
• Complete elimination: Removed objects vanish entirely from the scene
• Multiple criteria types: Tests different forms of object selection logic

data/questions/object_subtraction_task/object_subtraction_00000000/
├── first_frame.png      # Initial state (all objects present)
├── final_frame.png      # Final state (specified objects removed)
├── prompt.txt           # Task instructions with selection criteria
├── ground_truth.mp4     # Solution video (16 fps)
└── question_metadata.json # Task metadata

File specifications: Images are 1024×1024 PNG. Videos are MP4 at 16 fps, duration varies by number of objects removed.

object-manipulation selective-removal attribute-matching spatial-selection outlier-detection object-identification scene-editing