A FiftyOne Model Zoo integration for Isaac-0.2 by Perceptron AI - hybrid-reasoning vision-language models designed for real-world visual understanding tasks.
Isaac-0.2 extends the efficient frontier of perception — small models that outperform systems 10× larger on visual reasoning and perception tasks, all running on commodity GPUs or edge devices. From robotics to media search to industrial inspection, Isaac 0.2 delivers high-accuracy perception without the heavy compute footprint.
Read the full announcement.
Available Models:
- Isaac-0.2-2B-Preview - 2B parameter hybrid-reasoning model for maximum accuracy
- Isaac-0.2-1B - 1B parameter model for faster inference and edge deployment
- Reasoning via Thinking Traces: Short, structured reasoning traces improve multi-step decisions, small-object understanding, and ambiguous spatial tasks
- Tool Calling + Focus (Zoom & Crop): Isaac 0.2 can trigger tool calls to focus (zoom + crop) and re-query on smaller regions — improving fine-grained perception
- Structured Outputs: More reliable structured output generation for consistent JSON and predictable downstream integration
- Complex OCR: Improved text recognition across cluttered, low-resolution, or distorted regions — enabling accurate extraction from documents, diagrams, labels, screens, and dense real-world scenes
- Desktop Use: Better performance on everyday desktop and mobile workflows such as UI understanding and navigation, making Isaac faster and more capable for agentic use cases
- Object Detection: Detect and localize objects with bounding boxes
- Keypoint Detection: Identify key points in images with spatial awareness
- Complex OCR: Extract and detect text from documents, diagrams, labels, screens, and cluttered scenes
- Classification: Classify images into categories with reliable JSON output
- Visual Question Answering: Answer questions about image content
- Segmentation: Generate polygon masks for instance segmentation
- Desktop/UI Understanding: Navigate and understand desktop and mobile interfaces
pip install fiftyone
pip install perceptron
pip install transformers
pip install accelerate
pip install torch torchvision
pip install huggingface-hubimport fiftyone.zoo as foz
# Register this model zoo source
foz.register_zoo_model_source(
"https://github.com/perceptron-ai-inc/fiftyone-isaac-0_2",
overwrite=True
)import fiftyone.zoo as foz
# Load the Isaac-0.2 2B model
model = foz.load_zoo_model("PerceptronAI/Isaac-0.2-2B-Preview")
# Or load the 1B model for faster inference
model = foz.load_zoo_model("PerceptronAI/Isaac-0.2-1B")import fiftyone as fo
import fiftyone.zoo as foz
# Load a dataset
dataset = foz.load_zoo_dataset("quickstart", max_samples=10)
# Load model and set operation
model = foz.load_zoo_model("PerceptronAI/Isaac-0.2-2B-Preview")
model.operation = "detect"
# Set detection prompt
model.prompt = "Animals, Humans, Vehicles, Objects"
# Apply model to dataset
dataset.apply_model(model, label_field="detections")model.operation = "vqa"
model.prompt = "Describe the spatial relationships between objects in this scene"
dataset.apply_model(model, label_field="vqa_response")model.operation = "ocr_detection"
model.prompt = "Detect all text in this image"
dataset.apply_model(model, label_field="text_detections")model.operation = "ocr"
model.prompt = "Extract all text from this image"
dataset.apply_model(model, label_field="extracted_text")model.operation = "point"
model.prompt = "Identify key features: eyes, nose, corners"
dataset.apply_model(model, label_field="keypoints")model.operation = "classify"
model.prompt = "Classify the weather: sunny, rainy, snowy, cloudy, indoor"
dataset.apply_model(model, label_field="weather")model.operation = "segment"
model.prompt = "Draw polygons around the following objects: person, car, animal"
dataset.apply_model(model, label_field="polygons")Enable structured reasoning traces for improved accuracy on complex scenes. Thinking traces improve multi-step decisions, small-object understanding, and ambiguous spatial tasks:
model.operation = "detect"
model.enable_thinking = True
dataset.apply_model(model, label_field="detections_with_reasoning")
# Disable when done
model.enable_thinking = FalseEnable the Focus system (zoom + crop) for fine-grained perception. Isaac 0.2 can trigger tool calls to focus on smaller regions and re-query — improving detection of small objects and dense scenes. Only works with BOX operations (detect, ocr_detection):
model.operation = "detect"
model.enable_focus_tool_call = True
dataset.apply_model(model, label_field="focused_detections")
# Disable when done
model.enable_focus_tool_call = FalseYou can combine both options for maximum precision:
model.operation = "detect"
model.enable_thinking = True
model.enable_focus_tool_call = True
dataset.apply_model(model, label_field="enhanced_detections")You can use different prompts for each sample in your dataset:
# Apply model using the prompt field
model.operation = "detect"
dataset.apply_model(
model,
label_field="custom_detections",
prompt_field="sample_prompt"
)You can customize the system prompt for specific use cases:
model.system_prompt = """
You are a specialized assistant for medical image analysis.
Focus on identifying anatomical structures and abnormalities.
"""| Operation | Description | Output Type | Example Prompt |
|---|---|---|---|
detect |
Object detection with bounding boxes | fo.Detections |
"Cars, pedestrians, traffic signs" |
point |
Keypoint detection | fo.Keypoints |
"Eyes, nose, mouth corners" |
classify |
Image classification | fo.Classifications |
"Indoor or outdoor scene" |
ocr |
Text extraction | String | "Extract all text from the image" |
ocr_detection |
Text detection with boxes | fo.Detections |
"Detect text regions" |
ocr_polygon |
Text detection with polygons | fo.Polylines |
"Detect text regions" |
segment |
Instance segmentation | fo.Polylines |
"Segment all objects" |
vqa |
Visual question answering | String | "What is the main subject?" |
See isaac_0_2_demo.ipynb for a complete interactive notebook demonstrating all operations. You can run it directly in Google Colab or download it for local execution.
- Parameters: 2B (Preview) / 1B
- Architecture: Based on Qwen with custom vision encoder
- Vision Resolution: Dynamic, up to 60 megapixels
- Context Length: 16,384 tokens
- Training: Perceptive-language pretraining on multimodal data
- Code: Apache 2.0 License
- Model Weights: Creative Commons Attribution-NonCommercial 4.0 International License
- Introducing Isaac 0.2 (Blog Post)
- Isaac-0.2-2B-Preview on Hugging Face
- Isaac-0.2-1B on Hugging Face
- Try Isaac Demo
- Isaac API
- Perceptron SDK
- FiftyOne Documentation
If you use Isaac-0.2 in your research or applications, please cite:
@software{isaac2025fiftyone,
title = {Isaac-0.2 FiftyOne Model Zoo Integration},
author = {Perceptron AI},
year = {2025},
url = {https://github.com/perceptron-ai-inc/fiftyone-isaac-0_2},
note = {FiftyOne integration for Isaac-0.2 perceptive-language model}
}
@misc{perceptronai2025isaac,
title = {Isaac-0.2: A Perceptive-Language Model},
author = {{Perceptron AI}},
year = {2025},
publisher = {Hugging Face},
url = {https://huggingface.co/PerceptronAI/Isaac-0.2-2B-Preview},
note = {Open-source multimodal model for real-world visual understanding}
}- Technical inquiries: support@perceptron.inc
- Commercial inquiries: sales@perceptron.inc
- Join the team: join-us@perceptron.inc
- Perceptron AI for developing Isaac-0.2
- FiftyOne by Voxel51 for the model zoo framework