from transformers import Qwen3VLProcessor
from PIL import Image

# Load processor with custom max_pixels
processor = Qwen3VLProcessor.from_pretrained(
    'Qwen/Qwen3-VL-2B-Instruct',
    trust_remote_code=True,
    max_pixels=200_000  # Expect images to be resized to ~200k pixels
)

# Check the internal state
print(f"max_pixels attribute: {processor.image_processor.max_pixels}")
print(f"size['longest_edge']: {processor.image_processor.size['longest_edge']}")

# Process a 2000×2000 image (4M pixels)
test_image = Image.new('RGB', (2000, 2000), color='red')
print(f"Input image: {test_image.size[0]}×{test_image.size[1]} = {test_image.size[0] * test_image.size[1]:,} pixels")

# Further investigations
messages = [{"role": "user", "content": [
    {"type": "image", "image": test_image},
    {"type": "text", "text": "Describe this image."}
]}]

text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
result = processor(text=text, images=[test_image], return_tensors="pt")

# Check actual processed dimensions
grid_thw = result['image_grid_thw'][0]
temporal, height_patches, width_patches = grid_thw
total_patches = (temporal * height_patches * width_patches).item()
effective_pixels = total_patches * 16 * 16  # Each patch is 16×16 pixels

print(f"Grid (T, H, W): {temporal}×{height_patches}×{width_patches}")
print(f"Total patches: {total_patches}")
print(f"Effective pixels: {effective_pixels:,}")