layout-heron-GGUF

RT-DETRv2 Document Layout Detection (GGUF)

Document layout analysis model for CrispEmbed. Detects 17 region types in document images.

Architecture

• Backbone: ResNet-50-D (BN-folded Conv2d)
• Encoder: HybridEncoder (AIFI self-attention + FPN/PAN with CSP-RepVGG)
• Decoder: 6-layer transformer with deformable multi-scale cross-attention (300 queries)
• Classes: 17 (text, title, table, figure, formula, caption, section_header, list_item, footnote, page_header, page_footer, code, document_index, checkbox_selected, checkbox_unselected, form, key_value_region)
• Parameters: 42M
• Source: docling-project/docling-layout-heron (Apache-2.0)

Variants

Usage

# CLI
./build/crispembed -m layout-heron --layout document.png --json

# Server
./build/crispembed-server --layout layout-heron-q8_0.gguf
curl -X POST http://localhost:8080/layout/detect -d '{"image": "page.png"}'

from crispembed import CrispLayout
layout = CrispLayout("layout-heron-q8_0.gguf")
regions = layout.detect("document.png")

Parity

• Encoder: all stages cos=1.0 vs HF reference (with exact input)
• Detection score: 0.934 (HF reference: 0.955)
• 14 parity bugs found and fixed via systematic layer-by-layer diff

License

Apache-2.0 (same as upstream docling-layout-heron).

Provenance and EU AI Act Art. 53 note

• Upstream model: docling-project/docling-layout-heron — published by docling-project.
• Upstream licence: apache-2.0. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• What was done here: format conversion and/or quantisation only (GGUF). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository.
• Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.