pip install "pipecat-ai[moondream]"

from pipecat.services.moondream.vision import MoondreamService
from pipecat.frames.frames import VisionImageRawFrame
from PIL import Image

# Configure service
service = MoondreamService(
    model="vikhyatk/moondream2",
    revision="2024-08-26"
)

# Create pipeline
pipeline = Pipeline([
    image_input,      # Produces VisionImageRawFrame
    service,          # Analyzes images
    text_handler      # Handles text responses
])

# Example frame processing
image = Image.open("example.jpg")
frame = VisionImageRawFrame(
    format=image.mode,
    size=image.size,
    image=image.tobytes(),
    text="What objects are in this image?"
)

# Automatically uses CUDA if available
service = MoondreamService()

# Requires intel_extension_for_pytorch
import intel_extension_for_pytorch
service = MoondreamService()

service = MoondreamService(use_cpu=True)

# Optimize image before processing
def prepare_image(image_path):
    image = Image.open(image_path)
    # Resize if needed
    if max(image.size) > 1024:
        image.thumbnail((1024, 1024))
    return image

try:
    async for frame in service.run_vision(vision_frame):
        if isinstance(frame, ErrorFrame):
            logger.error(f"Vision processing error: {frame.error}")
        elif isinstance(frame, TextFrame):
            process_result(frame.text)
except Exception as e:
    logger.error(f"Unexpected error: {e}")

# Initialize once, reuse for multiple images
service = MoondreamService()
try:
    # Process multiple images
    for image in images:
        await process_image(service, image)
finally:
    # Cleanup if needed
    await service.cleanup()