Bug

When using the VLM pipeline (ibm-granite/granite-docling-258M via vLLM==0.11.2) with docling==2.64.0, the model occasionally predicts "flat" or marginal bounding boxes. When these normalized coordinates are converted to integer pixel coordinates by docling, they can result in a zero-height or zero-width crop box due to integer rounding.

This causes PIL to crash with SystemError: tile cannot extend outside image when ImageRef.from_pil attempts to save/process the crop.

Specific Edge Case Observed:
The VLM predicted: {'l': 0.398, 't': 0.998, 'r': 0.606, 'b': 0.998}.
Note that top and bottom are identical (0.998).

Even if a user manually adjusts bottom to 0.999 (a 0.001 difference), docling's internal logic converts these to pixels using int(). For a generic image height (e.g., 500px):

• Top: int(0.998 * 500) = 499
• Bottom: int(0.999 * 500) = 499

Result: top (499) == bottom (499). The crop height is 0 pixels, causing the crash.

Traceback:

File .../docling_core/types/doc/document.py:5670, in DoclingDocument.load_from_doctags
    image=ImageRef.from_pil(image=cropped_image, dpi=72),

File .../docling_core/types/doc/document.py:1127, in ImageRef.from_pil
    image.save(buffered, format="PNG")

... (PIL Internal calls) ...

File .../PIL/ImageFile.py:562, in _encode_tile
    encoder.setimage(im.im, extents)
SystemError: tile cannot extend outside image

Steps to reproduce

1. Environment:

   • docling: 2.64.0
   • vllm: 0.11.2
   • Model: ibm-granite/granite-docling-258M

2. Reproduction Script (Simulating the logic):
   Since the VLM output is non-deterministic, this script replicates the exact logic inside load_from_doctags that leads to the crash using the observed coordinates.

from PIL import Image
import io

# 1. Create a dummy image
im_width, im_height = 500, 500
image = Image.new("RGB", (im_width, im_height))

# 2. The problematic Bbox from VLM (Normalized)
# Top and Bottom are technically different floats, but very close.
bbox = {'l': 0.398, 't': 0.998, 'r': 0.606, 'b': 0.999} 

# 3. Docling's internal logic (converting to pixels)
crop_box = (
    int(bbox['l'] * im_width),
    int(bbox['t'] * im_height),
    int(bbox['r'] * im_width),
    int(bbox['b'] * im_height),
)

print(f"Calculated Crop Box: {crop_box}")
# Output: (199, 499, 303, 499) -> Height is 0!

# 4. Trigger the crash
# The crash happens here, when encoding the 0-height image to PNG
try:
    cropped = image.crop(crop_box)
    
    # Docling tries to save it to memory to create a Base64 string
    buf = io.BytesIO()
    cropped.save(buf, format="PNG") 
    print("Success (This line will not be reached)")
except Exception as e:
    print(f"\nCRASHED AS EXPECTED: {e}")

Suggested Fix:
The BoundingBox processing logic (or docling_core) needs to enforce a minimum pixel thickness after the int() conversion and before image.crop(). Validating float epsilon is insufficient because int(x) rounding can still collapse the dimension to zero.

Docling version

Docling version: 2.64.0
Docling Core version: 2.51.1
Docling IBM Models version: 3.10.0
Docling Parse version: 4.7.1
Python: cpython-310 (3.10.11)
Platform: macOS-26.1-arm64-arm-64bit

Python version

Python 3.10.11