MiliLab/Omni-I2C

---
license: apache-2.0
---

# Omni-I2C: Image2Code_Full

## Dataset Description

`Image2Code_Full.tsv` is the inference split of the **Omni-I2C** benchmark. It is designed to evaluate whether multimodal models can generate high-fidelity code or structured outputs from input images.

Each sample contains an image, an instruction, and metadata describing the target task. The goal is to generate code or a structured string that can reproduce the original figure as accurately as possible.

- **Number of samples:** 1,080
- **Number of subjects:** 8
- **Number of figure types:** 45
- **Number of code types:** 5

Project repository: [Omni-I2C on GitHub](https://github.com/MiliLab/Omni-I2C)

## Data Fields

The dataset is stored as a 7-column TSV file with the following header:

```tsv
index image question answer subject figure_type code_type
````

| Field | Description |
| ------------- | ------------------------------------------------------------------------------- |
| `index` | Unique sample identifier. |
| `image` | Input image stored as a base64-encoded string. |
| `question` | Instruction text for the generation task. |
| `answer` | Filename of the ground-truth reference file. This is not the full code content. |
| `subject` | Subject or application domain. |
| `figure_type` | Fine-grained figure category. |
| `code_type` | Target output format or code type. |

## Task Description

This is an **image-to-code** inference dataset. Given an input image and an instruction, the model is expected to generate an executable, renderable, or otherwise structured output that reconstructs the original figure.

The `answer` field links each sample to its corresponding ground-truth file used in evaluation.

## Supported Code Types

```text
python
html-css
latex-tikz
svg
smiles
```

### Sample Counts by Code Type

| code_type | Count | Extension |
| ------------ | ----: | --------- |
| `python` | 357 | `.py` |
| `latex-tikz` | 265 | `.tex` |
| `svg` | 192 | `.svg` |
| `html-css` | 166 | `.html` |
| `smiles` | 100 | `.smi` |

## Subjects

```text
Biology&Medicine
Chemistry
Computer Science
Economics
Geography
Math
Other
Physics
```

## Figure Types

This split includes 45 figure types, including but not limited to:

```text
3d-plot, Area, Contour, Density, Graph, Histogram, Phase-Diagram,
Quiver, Treemap, UML-class-diagram, Violin, analytical-geometry,
anatomy-diagram, atom-model, bar-chart, block-diagram, box-chart,
cell-structure, circuit, equations-texts, error-bar, error-point,
flow-chart, free-body-diagram, function-related, gauge-chart, graph,
heatmap, line-graph, magnetic-field-line, map, molecular-formula,
multi-graph, optics-ray-diagram, other-figures, physiological-process,
pie-chart, plane-geometry, radar-chart, relationship-diagram,
scatter-plot, schematic, solid-geometry, tables, venn-diagram
```

## Usage

Within the Omni-I2C project, this file is used as the inference input to the evaluation pipeline:

1. `VLMEvalKit_infer` loads `Image2Code_Full.tsv`
2. The model takes `image` and `question` as input
3. Predictions are saved after inference
4. `eval_pipeline` matches predictions with GT files for code-level and image-level evaluation

For implementation details, please refer to the project repository: [https://github.com/MiliLab/Omni-I2C](https://github.com/MiliLab/Omni-I2C)

## Citation

If you find Omni-I2C helpful, please consider citing the following paper:

```
@inproceedings{Zhou2026OmniI2CAH,
title={Omni-I2C: A Holistic Benchmark for High-Fidelity Image-to-Code Generation},
author={Jiawei Zhou and Chi Zhang and Xiang Feng and Qiming Zhang and Haibo Qiu and Lihuo He and Dengpan Ye and Xinbo Gao and Jing Zhang},
year={2026},
url={https://api.semanticscholar.org/CorpusID:286643606}
}
```