Does VlmPipeline ignore pipeline_options.do_picture_description = True?

[adrianad]

Wanted to use granite docling for OCR and Qwen3 VL for image description, but the option doesnt seem to work with the VlmPipeline. Nothing happens when I add it. I do have PictureItems in my result, but only with the caption. The Qwen 3 VL is never called. The models are running and i can see their VLLM logs, granite receives requests, Qwen doesnt.

Is this intended, or am I doing something wrong?

I tried to combine two examples from the docs:

import json
import logging
import os
from pathlib import Path
from typing import Optional

import requests
from docling_core.types.doc.page import SegmentedPage
from docling_core.types.doc import PictureItem
from dotenv import load_dotenv

from docling.datamodel.base_models import InputFormat
from docling.datamodel.pipeline_options import (
    VlmPipelineOptions,
    PictureDescriptionApiOptions
)
from docling.datamodel.pipeline_options_vlm_model import ApiVlmOptions, ResponseFormat
from docling.document_converter import DocumentConverter, PdfFormatOption
from docling.pipeline.vlm_pipeline import VlmPipeline

### Example of ApiVlmOptions definitions

#### Using LM Studio or VLLM (OpenAI-compatible APIs)


def openai_compatible_vlm_options(
    model: str,
    prompt: str,
    format: ResponseFormat,
    hostname_and_port,
    temperature: float = 0.7,
    max_tokens: int = 4096,
    api_key: str = "",
    skip_special_tokens=False,
):
    headers = {}
    if api_key:
        headers["Authorization"] = f"Bearer {api_key}"

    options = ApiVlmOptions(
        url=f"http://{hostname_and_port}/v1/chat/completions",  # LM studio defaults to port 1234, VLLM to 8000
        params=dict(
            model=model,
            max_tokens=max_tokens,
            skip_special_tokens=skip_special_tokens,  # needed for VLLM
        ),
        headers=headers,
        prompt=prompt,
        timeout=90,
        scale=2.0,
        temperature=temperature,
        response_format=format,
        concurrency=20,
    )
    return options

def vllm_local_options(model: str):
    options = PictureDescriptionApiOptions(
        url="http://x.x.x.x:8000/v1/chat/completions",
        params=dict(
            model=model,
            seed=42,
            max_completion_tokens=200,
        ),
        prompt="Describe the image.",
        timeout=90,
        concurrency=20,
    )
    return options



### Usage and conversion
logging.basicConfig(level=logging.DEBUG)

input_doc_path = "./Downloads/Open-ended_Deep_Reinforcement_Learning_for_a_Bitcoin_Trading_Bot.pdf"

# Configure the VLM pipeline. Enabling remote services allows HTTP calls to
# locally hosted APIs (LM Studio, Ollama) or cloud services.
pipeline_options = VlmPipelineOptions(
    enable_remote_services=True,  # required when calling remote VLM endpoints
    generate_picture_images=True,
    do_picture_description=True,
    picture_description_options=vllm_local_options(model="LLM")
)

# Example using the Granite-Docling model with LM Studio or VLLM:
pipeline_options.vlm_options = openai_compatible_vlm_options(
    model="ibm-granite/granite-docling-258M",  # For VLLM use "ibm-granite/granite-docling-258M"
    hostname_and_port="x.x.x.x:8001",  # LM studio defaults to port 1234, VLLM to 8000
    prompt="Convert this page to docling.",
    format=ResponseFormat.DOCTAGS,
    api_key="dummy",
)

# Create the DocumentConverter and launch the conversion.
doc_converter = DocumentConverter(
    format_options={
        InputFormat.PDF: PdfFormatOption(
            pipeline_options=pipeline_options,
            pipeline_cls=VlmPipeline,
        )
    }
)
result = doc_converter.convert(input_doc_path)
print(result.document.export_to_markdown())

[LifangD]

it seems the markdown format that works

[david-kort]

I'm experiencing the same issue; as soon as I switch from doclings standard OCR pipeline to the VlmPipeline, the enabled do_picture_description and do_picture_classification enrichment feature settings don't have any effect and are ignored.

[awschult002]

this has been my experience. the VLM pipeline seems to expect that a single pass through a VLM model will do EVERYTHING including extracting text and generating the descriptions or classifications for images/figures. However, when using Granite-Docling (default) that model is NOT designed to do general image descriptions, it only extracts text and labels images/figures.

The ONLY WAY i have gotten a more complete package is by taking the VLM response, then looping through the doctags for figures, then sending those figures off one at a time to another model to get descriptions. I even tried working with the enrichment pipelines to no avail.

This is incredibly annoying and I desperately need it to be fixed before I can deploy it for my company. (VLM is a must, we have lots of scanned documents)