📚 Introduction

We present the InternSVG family, an integrated data–benchmark–model suite.

• 🧩 SAgoge Dataset — The largest and most comprehensive multimodal dataset for SVG tasks, spanning icons, long-sequence illustrations, scientific diagrams, and dynamic animations. It provides rich hierarchical structures and diverse attributes, supporting tasks of varied difficulty levels.
• 📊 SArena Benchmark — A companion benchmark offering unified task definitions and standardized evaluation protocols, aligned with SAgoge’s domains and difficulty spectrum. It enables consistent comparison across SVG understanding, editing, and generation tasks.
• 🤖 InternSVG Model — A unified multimodal large language model (MLLM) for SVG understanding, editing, and generation.

🔥 News

• [2026-01-28] 🎉 InternSVG-8B is now available on HuggingFace! 🤗Model
• [2026-01-28] 🎉 We release the SAgoge dataset. 🤗Dataset
• [2026-01-26] 🎉 InternSVG has been accepted at ICLR 2026!
• [2025-10-13] 🎉 We release the SArena benchmark. 🤗Benchmark
• [2025-10-13] 👋 Upload paper and init project. Read

📝 Open-Source Plan

• Evaluation code
• SArena benchmark
• SAgoge dataset
• Fine-tuning scripts
• Model weights
• Paper

📌 Quick Start

⚙️ Installation

git clone https://github.com/hmwang2002/InternSVG.git
cd InternSVG

conda create -n internsvg python=3.9 -y
conda activate internsvg
pip install -r requirements.txt

# install clip
pip install git+https://github.com/openai/CLIP.git

Download ViCLIP.

mkdir sarena_ckpt
cd sarena_ckpt
# You need to login first and have the access to the repo https://huggingface.co/OpenGVLab/ViCLIP. Use the command "huggingface-cli login" to login.
huggingface-cli download --resume-download OpenGVLab/ViCLIP ViClip-InternVid-10M-FLT.pth --local-dir .
cd ..

For training, you need to install LLaMA-Factory.

pip install deepspeed==0.16.9
pip install av==14.4.0
cd LLaMA-Factory
pip install -e ".[torch,metrics]"
cd ..

(Optional) If you need to simplify your own SVG code, install svgo.

conda install nodejs
npm install -g svgo

🤖 InternSVG Model

The InternSVG-8B model is available at Hugging Face. It is based on the InternVL3-8B model, incorporating SVG-specific tokens, and undergoes Supervised Fine-Tuning (SFT) under a two-stage training strategy using the massive SVG training samples from the SAgoge dataset.

Deploy

We recommend using LMDeploy for deployment. An example of launching a proxy server with 8 parallel workers (one per GPU) is provided below:

#!/bin/bash
model_path="MODEL_PATH"
model_name="InternSVG"

# proxy
lmdeploy serve proxy --server-name 0.0.0.0 --server-port 10010 --routing-strategy "min_expected_latency" &

worker_num=8
for ((i = 0; i < worker_num; i++)); do
    timestamp=$(date +"%Y-%m-%d_%H-%M-%S")
    CUDA_VISIBLE_DEVICES="${i}" lmdeploy serve api_server ${model_path} --proxy-url http://0.0.0.0:10010 \
        --model-name ${model_name} \
        --tp 1 \
        --max-batch-size 512 \
        --backend pytorch \
        --server-port $((10000 + i)) \
        --session-len 16384 \
        --chat-template "internvl2_5" \
        --log-level WARNING &>> ./logs/api_${model_name}_${timestamp}_${i}.out  &
    sleep 10s
done

Train

If you need to train your own model, please follow these steps:

1. Prepare the Dataset: Download the SAgoge dataset. After that, update the paths for the SAgoge-related subdatasets in LLaMA-Factory/data/dataset_info.json to match your local file paths.

2. Download InternVL3-8B: Download the InternVL3-8B from link.

3. Add Special Tokens: Before training, you must add SVG-specific tokens to the base model. Run the utils/add_token.py script, which adds these special tokens to the original model weights and initializes their embeddings based on subwords.

4. Start Training: We provide example configuration scripts for the two-stage training process. You can find them at:

   • Stage 1: LLaMA-Factory/examples/train_full/stage_1.yaml
   • Stage 2: LLaMA-Factory/examples/train_full/stage_2.yaml

   Then use llamafactory-cli train to start training.

🧩 SAgoge Dataset

The SAgoge dataset is available at Hugging Face. To use SAgoge, please download the dataset and extract media.tar.gz to access the image files. After extraction, you will get:

SAgoge/
├── media/
│   ├── stage1/
│   │   ├── chem/
│   │   └── icon/
│   └── stage2/
│       ├── animation/
│       ├── chem/
│       ├── icon/
│       └── illustration/
├── stage1/
│   ├── chem/
│   │   ├── img2svg/
│   │   └── text2svg/
│   └── icon/
│       ├── edit/
│       ├── generation/
│       │   ├── img2svg/
│       │   └── text2svg/
│       └── understanding/
└── stage2/
    ├── animation/
    │   ├── text2sani/
    │   └── video2sani/
    ├── chem/
    │   ├── img2svg/
    │   └── text2svg/
    ├── icon/
    │   ├── edit/
    │   ├── generation/
    │   │   ├── img2svg/
    │   │   └── text2svg/
    │   └── understanding/
    └── illustration/
        ├── img2svg/
        └── text2svg/

Statistics of SAgoge:

📊 SArena Benchmark

Download

The SArena benchmark is available here. You can use the huggingface_hub command to download directly:

hf download InternSVG/SArena SArena.zip --repo-type dataset --resume-download --local-dir PATH_TO_YOUR_DIR
unzip SArena.zip

After extraction, you will get:

SArena/
├── animation/
│   ├── overall/
│   ├── svg/
│   ├── video/
│   ├── text2sani.jsonl
│   └── video2sani.jsonl
│
├── chemistry/
│   ├── images/
│   ├── svg/
│   ├── img2svg.jsonl
│   └── text2svg.jsonl
│
├── illustration/
│   ├── images/
│   ├── svg/
│   ├── caption.jsonl
│   ├── img2svg.jsonl
│   └── text2svg.jsonl
│
├── Icon/
│   ├── edit/
│   │   └── data/
│   │       ├── color_complex.jsonl
│   │       ├── color_simple.jsonl
│   │       ├── crop.jsonl
│   │       ├── flip.jsonl
│   │       ├── opacity.jsonl
│   │       ├── outline.jsonl
│   │       ├── rotate.jsonl
│   │       ├── scale.jsonl
│   │       ├── styletransform_openmoji.jsonl
│   │       └── translate.jsonl
│   │
│   ├── generation/
│   │   ├── images/
│   │   ├── svg/
│   │   ├── caption.jsonl
│   │   ├── img2svg.jsonl
│   │   └── text2svg.jsonl
│   │
│   └── understanding/
│       └── sarena_un.jsonl

Inference

Template scripts for inference can be found in the scripts/inference/ folder.

For example, for the icon/illustration/chemistry generation task, you can modify the script above by specifying your own paths and API configuration.

#!/bin/bash
export PYTHONPATH=$(pwd):$PYTHONPATH

BASE_URL="BASE_URL"
API_KEY="API_KEY"
MODEL_NAME="MODEL_NAME"
TEXT2SVG_TEST_PATH="PATH_TO_TEXT2SVG_TEST_PATH"
IMG2SVG_TEST_PATH="PATH_TO_IMG2SVG_TEST_PATH"
OUTPUT_DIR="PATH_TO_OUTPUT_DIR"
RETRY=1
TEMPERATURE=0.0
MAX_TOKENS=4000
MAX_WORKERS=32

python metrics/inference/inference.py \
--base_url ${BASE_URL} \
--api_key ${API_KEY} \
--model_name ${MODEL_NAME} \
--text2svg_test_path ${TEXT2SVG_TEST_PATH} \
--img2svg_test_path ${IMG2SVG_TEST_PATH} \
--output_dir ${OUTPUT_DIR} \
--temperature ${TEMPERATURE} \
--max_tokens ${MAX_TOKENS} \
--max_workers ${MAX_WORKERS}

Then run:

bash scripts/inference/gen/demo.sh

Specifically, for SVG animation generation task, a template inference script is provided at scripts/inference/animation/demo.sh.

When all test samples have been processed, each SVG file needs to be converted into an MP4 video for metric evaluation. Use the script utils/svg_animate.py to generate MP4 files. Note that we need two resolutions: 448×448 and 128×128. Before running, modify the OUTPUT_DIRS and FILE_DIRS variables in the run_all_mp() function. (Notably, in our code, if the output path contains '_128', it will automatically use the 128×128 resolution.)

The directory structure of the test files is as follows:

evaluate
├── .vscode
├── animation/gpt4o
│   ├── text2sani
│   │   ├── svg/
│   │   ├── video/
│   │   ├── video_128/
│   │   └── output.jsonl
│   └── video2sani
│       ├── svg/
│       ├── video/
│       ├── video_128/
│       └── output.jsonl

Evaluate

The scripts/evaluate/ directory contains template scripts for running evaluation across different domains (e.g., icon, illustration, chemistry, and animation).

Each subfolder corresponds to a specific domain:

scripts/evaluate/
├── icon/
│   ├── edit/
│   ├── gen/
│   └── un/
├── illustration/
├── chem/
└── animation/

Below is a demo for evaluating generation tasks (Text-to-SVG and Image-to-SVG):

#!/bin/bash
export PYTHONPATH=$(pwd):$PYTHONPATH

python evaluate_gen.py \
    --model_name "GPT-4o" \
    --text2svg_test_dir "PATH_TO_TEXT2SVG_RESULTS" \
    --img2svg_test_dir "PATH_TO_IMG2SVG_RESULTS" \
    --tokenizer_path "PATH_TO_TOKENIZER" \
    --test_file_path "PATH_TO_TEST_JSONL" \
    --gt_img_dir "PATH_TO_GT_IMAGES" \
    --gt_svg_dir "PATH_TO_GT_SVGS" \
    --caption_path "PATH_TO_CAPTIONS" \
    --bench_name "Icon"

If your model does not support either the Text-to-SVG or Image-to-SVG task, simply set the corresponding test directory argument (--text2svg_test_dir or --img2svg_test_dir) to an empty string.

Note on the SArena-Icon Score Reported by Intern-S1-Pro

We thank the Intern-S1-Pro team for treating SVG generation as an important capability of the model, and for using SArena-Icon as a benchmark for evaluation.

To better align with the common practice in general-purpose foundation model technical reports (i.e., reporting results in a 0 – 100 scale), we additionally report a single aggregated score for SArena-Icon, computed from the original SArena metrics.

We define the final score as:

final score = 0.3 * CLIP-I2I + 0.3 * (100 * DINO) + 0.2 * (100 * SSIM) + 0.2 * (100 * (1 - LPIPS))

📜 Acknowledgements

We would like to thank Kiyotaka, yinlikestudy, and quentin-77 for their valuable contributions to this project.

The InternSVG model is developed based on InternVL and further fine-tuned with LLaMA-Factory for SVG understanding, editing, and generation tasks.

We also acknowledge the following open-source efforts that have contributed to advancing SVG understanding and generation:

• SGP-Bench
• StarVector
• LLM4SVG
• OmniSVG

License

InternSVG is licensed under the Apache License 2.0.

📖 Citation

@inproceedings{wang2026internsvg,
  title={Internsvg: Towards unified svg tasks with multimodal large language models},
  author={Wang, Haomin and Yin, Jinhui and Wei, Qi and Zeng, Wenguang and Gu, Lixin and Ye, Shenglong and Gao, Zhangwei and Wang, Yaohui and Zhang, Yanting and Li, Yuanqi and others},
  booktitle={International Conference on Learning Representations},
  volume={2026},
  pages={38649--38705},
  year={2026}
}