---
license: apache-2.0
library_name: transformers
tags:
  - multimodal
  - vision-language
  - code-generation
  - tikz
  - geometric-reasoning
  - computer-vision
  - cvpr2026
  - internvl
  - internlm2
datasets:
  - SJY-1995/GeoTikz-Base
model-index:
  - name: GeoTikzBridge-Base-8B
    results:
      - task:
          type: image-to-text
          name: Image-to-TikZ Code Generation
        dataset:
          name: GeoTikz-Base
          type: SJY-1995/GeoTikz-Base
        metrics:
          - type: CLIP-S
            value: 89.5
            name: CLIP-S
          - type: compilation_accuracy
            value: 97.1
            name: LaTeX Compilation Success Rate
---

# GeoTikzBridge-Base-8B

## Model Overview
GeoTikzBridge-Base-8B is a multimodal large language model proposed in the **CVPR 2026** accepted paper *GeoTikzBridge: Advancing Multimodal Code Generation for Geometric Perception and Reasoning*, specializing in geometric figure perception and end-to-end TikZ code generation. Built on the InternVL multimodal architecture, this model is fully fine-tuned on the 2.5M-scale GeoTikz-Base dataset, enabling accurate conversion of geometric images into directly compilable, standard-compliant TikZ code, with end-to-end optimization for fine-grained geometric perception, spatial relationship reasoning, and structured code generation.

## Model Details
### Core Architecture
- Backbone Foundation: Built on the InternVL2 8B (InternLM2 7B backbone) multimodal large language model, inheriting the state-of-the-art vision-language alignment capability and code generation proficiency of the InternVL series.
- Parameter Scale: 8 billion parameters, balancing inference efficiency and generation accuracy, supporting deployment and implementation on consumer-grade GPUs.
- Targeted Optimization: Specialized alignment for fine-grained features of geometric figures (including lines, angles, annotations, coordinate relationships, etc.), significantly improving the restoration accuracy of geometric structures and the compilability of generated TikZ code.

### Core Capabilities
1.  High-precision geometric image-to-TikZ code conversion, with output code directly renderable in LaTeX environments to generate vector geometric figures highly consistent with the input image.
2.  Fine-grained geometric element perception, supporting complete restoration of complex planar geometric figures including points, lines, planes, angle annotations, dimension labels, and nested mathematical formulas.
3.  Standardized code output, with generated TikZ code complying with LaTeX typesetting specifications, featuring clear structure and easy secondary modification.

## Intended Use & Limitations
### Intended Use Cases
- Core Scenarios: Vectorization of geometric illustrations for academic papers, generation of geometric materials for textbooks/courseware, auxiliary tools for geometry teaching, and codified conversion of planar engineering drawings.
- Research Purposes: Serves as a baseline model for research in geometric perception and multimodal code generation, supporting secondary development and academic research in related fields.
- Downstream Expansion: Can be further fine-tuned on this model to expand advanced capabilities such as instruction following, auxiliary line generation, and geometry problem solving (corresponding to the Instruct series models of this project).

### Out-of-Scope Use Cases
- Code generation for non-geometric images (e.g., natural landscapes, portraits, complex 3D modeling, unstructured hand-drawn doodles, etc.).
- Generation of extremely complex 3D geometric figures and high-precision engineering assembly drawings beyond the capability of TikZ planar drawing.
- Drawing generation for high-risk industrial/engineering scenarios without compilation and validation.

### Model Limitations
- The TikZ code generated by the model requires compilation and validation via LaTeX; locally structural deviations may occur for extremely complex composite geometric figures.
- The model only supports annotation generation in English and mathematical symbols, and does not support text annotation content in other languages for now.
- Generation performance is significantly affected by the clarity and standardization of the input image; blurry, distorted, or severely occluded input images may lead to degraded generation quality.

## Quick Start
### Environment Setup
Install basic dependencies:
```bash
pip install transformers torch pillow accelerate
```
For full training/inference dependencies, please refer to the official project repository: [GeoTikzBridge GitHub](sslocal://flow/file_open?url=https%3A%2F%2Fgithub.com%2Fsjy-1995%2FGeoTikzBridge%2F&flow_extra=eyJsaW5rX3R5cGUiOiJjb2RlX2ludGVycHJldGVyIn0=)

### Inference Example
Quickly load the model with the Transformers library to achieve end-to-end image-to-TikZ code generation:
```python
from transformers import AutoProcessor, AutoModelForCausalLM
import torch
from PIL import Image

# Load model and processor
model_name = "SJY-1995/GeoTikzBridge-Base-8B"
processor = AutoProcessor.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
    device_map="auto"
)

# Load input geometric image
image = Image.open("your_geometric_figure.png").convert("RGB")

# Build inference prompt
prompt = ""
inputs = processor(text=prompt, images=image, return_tensors="pt").to(model.device)

# Generate TikZ code
with torch.no_grad():
    output = model.generate(
        **inputs,
        max_new_tokens=4096,
        temperature=0.1,
        top_p=0.95,
        do_sample=False
    )

# Decode and output the generated result
tikz_code = processor.decode(output[0], skip_special_tokens=True)
print("Generated TikZ Code:\n", tikz_code)
```

## Training Details
### Training Dataset
The model is fully fine-tuned on the **GeoTikz-Base** dataset, which contains approximately 2.5 million high-quality geometric image-TikZ code paired samples. The dataset covers full scenarios including basic planar geometric figures, composite figures, nested annotations, and formula integration, building a precise vision-code aligned corpus.

Dataset Link: [SJY-1995/GeoTikz-Base](sslocal://flow/file_open?url=https%3A%2F%2Fhuggingface.co%2Fdatasets%2FSJY-1995%2FGeoTikz-Base&flow_extra=eyJsaW5rX3R5cGUiOiJjb2RlX2ludGVycHJldGVyIn0=)

### Key Training Hyperparameters
| Hyperparameter | Configuration |
|----------------|---------------|
| Global Batch Size | 128 |
| Peak Learning Rate | 4e-7 |
| Training Epochs | 3 |
| Max Sequence Length | 12800 |
| Training Precision | BF16 |

### Training Framework & Scripts
Model training is implemented based on the official InternVL training framework. The core fine-tuning scripts and complete training pipeline can be found in the project repository:
- Core Script Path: `./internvl_chat/shell/internvl2.0/2nd_finetune/internvl2_8b_internlm2_7b_dynamic_res_2nd_finetune_full.sh`
- Project Repository: [GeoTikzBridge GitHub](sslocal://flow/file_open?url=https%3A%2F%2Fgithub.com%2Fsjy-1995%2FGeoTikzBridge%2F&flow_extra=eyJsaW5rX3R5cGUiOiJjb2RlX2ludGVycHJldGVyIn0=)

## Model Family
The GeoTikzBridge series includes multiple models with different specifications and capabilities to adapt to various scenario requirements:
| Model Name | Parameter Size | Core Capability | Model Link |
|------------|----------------|-----------------|------------|
| GeoTikzBridge-Base-8B | 8B | Basic geometric image-to-TikZ code generation | [🤗 Hugging Face](sslocal://flow/file_open?url=https%3A%2F%2Fhuggingface.co%2FSJY-1995%2FGeoTikzBridge-Base-8B&flow_extra=eyJsaW5rX3R5cGUiOiJjb2RlX2ludGVycHJldGVyIn0=) |
| GeoTikzBridge-Base-38B | 38B | High-precision TikZ code generation for complex geometric figures | [🤗 Hugging Face](sslocal://flow/file_open?url=https%3A%2F%2Fhuggingface.co%2FSJY-1995%2FGeoTikzBridge-Base-38B&flow_extra=eyJsaW5rX3R5cGUiOiJjb2RlX2ludGVycHJldGVyIn0=) |
| GeoTikzBridge-Instruct-8B | 8B | Instruction following, auxiliary line generation, interactive geometric reasoning | [🤗 Hugging Face](sslocal://flow/file_open?url=https%3A%2F%2Fhuggingface.co%2FSJY-1995%2FGeoTikzBridge-Instruct-8B&flow_extra=eyJsaW5rX3R5cGUiOiJjb2RlX2ludGVycHJldGVyIn0=) |

## Citation
If you use this model, related datasets or code in your research or projects, please cite the following paper:
```bibtex
@inproceedings{
  geotikzbridge,
  title={GeoTikzBridge: Advancing Multimodal Code Generation for Geometric Perception and Reasoning},
  author={Jiayin Sun and Caixia Sun and Boyu Yang and Hailin Li and Xiao Chen and Yi Zhang and Errui Ding and Liang Li and Chao Deng and Junlan Feng},
  booktitle={2026 IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2026}
}
```