Granite Docling extract images with higher resolution

[paoloaveri]

Bug

Using granite docling, I can't extract images with a good resolution.

Steps to reproduce

Wrote a simple script that works on page 5 only of https://arxiv.org/pdf/2501.17887:

from pathlib import Path
from docling.document_converter import DocumentConverter
from docling.datamodel.base_models import InputFormat
from docling.document_converter import PdfFormatOption
from docling_core.types.doc import ImageRefMode
from docling.datamodel.pipeline_options import VlmPipelineOptions
from docling.pipeline.vlm_pipeline import VlmPipeline
from docling.datamodel.vlm_model_specs import GRANITEDOCLING_TRANSFORMERS as granite_docling_vlm_conversion_options

output_dir = Path("output")
output_dir.mkdir(exist_ok=True)

docling_paper = "arvix.pdf" # Only page 5 of https://arxiv.org/pdf/2501.17887

vlm_options = granite_docling_vlm_conversion_options
IMAGE_RESOLUTION_SCALE = 1.0

vlm_pipeline_options = VlmPipelineOptions(
    force_backend_text=False,
    vlm_options=vlm_options,
    images_scale=IMAGE_RESOLUTION_SCALE,
    generate_picture_images=True
)

converter_vlm = DocumentConverter(
    format_options={
        InputFormat.PDF: PdfFormatOption(
            pipeline_options=vlm_pipeline_options,
            pipeline_cls=VlmPipeline,
        )
    }
)

vlm_result = converter_vlm.convert(docling_paper)
vlm_doc = vlm_result.document
html_filename = output_dir / f"{docling_paper}.html"
vlm_doc.save_as_html(html_filename, image_mode=ImageRefMode.REFERENCED)

With this, the images are low resolution (on top on being badly framed):

When pushing IMAGE_RESOLUTION_SCALE to 2.0, then we see this bug where the images are extracted from somewhere else in the document it looks like:

Docling version

2025-10-09 11:24:55,555 - INFO - Loading plugin 'docling_defaults'
2025-10-09 11:24:55,560 - INFO - Registered ocr engines: ['easyocr', 'ocrmac', 'rapidocr', 'tesserocr', 'tesseract']
Docling version: 2.55.1
Docling Core version: 2.48.4
Docling IBM Models version: 3.9.1
Docling Parse version: 4.5.0
Python: cpython-311 (3.11.10)
Platform: macOS-15.7.1-x86_64-i386-64bit

Python version

Python 3.11.10

[dosubot]

Hi @paoloaveri! I'm Dosu and I’m helping the docling team.

This looks like a bug in how Granite Docling handles image region scaling. When you increase IMAGE_RESOLUTION_SCALE, the pipeline scales both the page image and the bounding box used for cropping, but if the bounding box coordinates are already in the scaled image space (or not consistently in the original PDF space), this double-scaling causes the crop to select the wrong region—resulting in badly framed or incorrect images at higher resolutions. The fallback logic may also select unrelated embedded images if cropping fails.

There’s no documented workaround or config that reliably extracts high-res images without these errors. The root cause is likely a mismatch in coordinate conventions between the bounding box and the scaled image, and fixing it would require ensuring all bounding boxes are always in the original page coordinate space before scaling for cropping. You can see the relevant scaling and cropping logic in the pipeline code here.

If you need a temporary workaround, you could try extracting images at scale 1.0 and upscaling them manually, but that won’t improve framing. Otherwise, a code fix is needed to make the bounding box scaling robust.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}