import pathlib from PIL import Image from lis1_infer import LIS1Inferencer def main() -> None: # 1. Instantiate the inference class and enable debug logging inferencer = LIS1Inferencer() inferencer.debug = True # 2. Load Qwen3.5 backbone + LIS1 decoder + LoRA adapter (with 4-bit bitsandbytes quantization) inferencer.load_model( qwen_path="./Qwen3.5-9B", lis_model_path="./LIS1-9B-forLora.safetensors", lora_adapter="./LIS1-9B-LoraAdapter", device="cuda", max_pixels=524288, load_in_4bit=True, # Or set load_in_8bit=True for 8-bit quantization ) # 3. Load "image1.jpg" and define target queries pil_images = [Image.open("image1.jpg").convert("RGB")] queries_per_image = [["toast", "red sauce", "something you can drink"]] # 4. Run inference (isolate_queries=True avoids multi-query cross-bleed while reusing cached ViT features) outputs = inferencer.run_model( images=pil_images, queries=queries_per_image, threshold=0.7, batch_size=4, isolate_queries=True, # Compute queries separately cache_vision=True, # Cache vision when above is True for better performance exclusive_masks=True, # Make masks not overlap, choose the best one ) for img_idx, out in enumerate(outputs): print(f"\nImage {img_idx} ({out['image_size']['width']}x{out['image_size']['height']}):") for obj in out["objects"]: print( f" - [{obj['id']}] '{obj['description']}': " f"score={obj['score']:.4f}, bbox={obj['bbox']}, area={obj['area']}" ) # 5. Render masks and bounding boxes onto the image and save rendered_images = inferencer.render_image( images=pil_images, outputs=outputs, draw_masks=True, draw_boxes=True, draw_labels=True, draw_scores=True, draw_contours=True, ) out_dir = pathlib.Path(".") out_dir.mkdir(parents=True, exist_ok=True) for i, img_full in enumerate(rendered_images): full_path = out_dir / f"result_{i}_boxes_and_masks.png" img_full.save(full_path) print(f"Saved rendered output to: {full_path}") # 6. Offload model from GPU/CPU memory when done inferencer.offload_model() if __name__ == "__main__": main()