feat: Heron layout model as new default (#1971)

* feat: Switch default layout model to DOCLING_LAYOUT_HERON. Update the unit test data.

Signed-off-by: Nikos Livathinos <[email protected]>

* Use default layout model in model_downloader default args

Signed-off-by: Christoph Auer <[email protected]>

* Use default layout model in model_downloader default args

Signed-off-by: Christoph Auer <[email protected]>

* Update docling-models tag for TableFormer

Signed-off-by: Christoph Auer <[email protected]>

* Update test GT

Signed-off-by: Christoph Auer <[email protected]>

* Update test GT (from linux CPU)

Signed-off-by: Ubuntu <[email protected]>

* fix: Ensure that the visualisations happen on copies of the page image

Signed-off-by: Nikos Livathinos <[email protected]>

* chore: Pinpoint docling-ibm-models to the fix branch for the ReadingOrderPredictor

Signed-off-by: Nikos Livathinos <[email protected]>

* chore: Update uv.lock

Signed-off-by: Nikos Livathinos <[email protected]>

* chore: Update tests GT to match the Heron layout model and the improved reading order model in Linux

Signed-off-by: Nikos Livathinos <[email protected]>

* fix: Introduce the verify_doctags optional parameter in conversion tests to control if a doctags
comparison should take place. Skip doctags comparisons for certain tests.

Signed-off-by: Nikos Livathinos <[email protected]>

* chore: Generate tests GT on Mac

Signed-off-by: Nikos Livathinos <[email protected]>

* chore: Remove the pinning of the docling-ibm-models and use the release 3.9.1

Signed-off-by: Nikos Livathinos <[email protected]>

---------

Signed-off-by: Nikos Livathinos <[email protected]>
Signed-off-by: Christoph Auer <[email protected]>
Signed-off-by: Ubuntu <[email protected]>
Co-authored-by: Christoph Auer <[email protected]>
Co-authored-by: Ubuntu <[email protected]>

Author: 3 people

docling/datamodel/pipeline_options.py

@@ -283,10 +283,10 @@ class LayoutOptions(BaseModel):
     keep_empty_clusters: bool = (
         False  # Whether to keep clusters that contain no text cells
     )
+    model_spec: LayoutModelConfig = DOCLING_LAYOUT_HERON
     skip_cell_assignment: bool = (
         False  # Skip cell-to-cluster assignment for VLM-only processing
     )
-    model_spec: LayoutModelConfig = DOCLING_LAYOUT_V2
 
 
 class AsrPipelineOptions(PipelineOptions):

docling/models/layout_model.py

@@ -91,7 +91,7 @@ def download_models(
         local_dir: Optional[Path] = None,
         force: bool = False,
         progress: bool = False,
-        layout_model_config: LayoutModelConfig = DOCLING_LAYOUT_V2,
+        layout_model_config: LayoutModelConfig = LayoutOptions().model_spec,  # use default
     ) -> Path:
         return download_hf_model(
             repo_id=layout_model_config.repo_id,
@@ -122,8 +122,8 @@ def draw_clusters_and_cells_side_by_side(
         left_clusters = [c for c in clusters if c.label not in exclude_labels]
         right_clusters = [c for c in clusters if c.label in exclude_labels]
         # Create a deep copy of the original image for both sides
-        left_image = copy.deepcopy(page.image)
-        right_image = copy.deepcopy(page.image)
+        left_image = page.image.copy()
+        right_image = page.image.copy()
 
         # Draw clusters on both images
         draw_clusters(left_image, left_clusters, scale_x, scale_y)

docling/models/page_preprocessing_model.py

@@ -90,7 +90,7 @@ def _parse_page_cells(self, conv_res: ConversionResult, page: Page) -> Page:
 
         # DEBUG code:
         def draw_text_boxes(image, cells, show: bool = False):
-            draw = ImageDraw.Draw(image)
+            draw = ImageDraw.Draw(image.copy())
             for c in cells:
                 x0, y0, x1, y1 = (
                     c.to_bounding_box().l,

docling/models/table_structure_model.py

@@ -94,7 +94,7 @@ def download_models(
     ) -> Path:
         return download_hf_model(
             repo_id="ds4sd/docling-models",
-            revision="v2.2.0",
+            revision="v2.3.0",
             local_dir=local_dir,
             force=force,
             progress=progress,

docling/utils/model_downloader.py

@@ -4,6 +4,7 @@
 
 from docling.datamodel.layout_model_specs import DOCLING_LAYOUT_V2
 from docling.datamodel.pipeline_options import (
+    LayoutOptions,
     granite_picture_description,
     smolvlm_picture_description,
 )
@@ -47,7 +48,7 @@ def download_models(
     if with_layout:
         _log.info("Downloading layout model...")
         LayoutModel.download_models(
-            local_dir=output_dir / DOCLING_LAYOUT_V2.model_repo_folder,
+            local_dir=output_dir / LayoutOptions().model_spec.model_repo_folder,
             force=force,
             progress=progress,
         )

pyproject.toml

@@ -46,7 +46,7 @@ dependencies = [
   'pydantic (>=2.0.0,<3.0.0)',
   'docling-core[chunking] (>=2.42.0,<3.0.0)',
   'docling-parse (>=4.2.2,<5.0.0)',
-  "docling-ibm-models>=3.9.0,<4",
+  "docling-ibm-models>=3.9.1,<4",
   'filetype (>=1.2.0,<2.0.0)',
   'pypdfium2 (>=4.30.0,!=4.30.1,<5.0.0)',
   'pydantic-settings (>=2.3.0,<3.0.0)',