Update README.md

README.md

@@ -1,3 +1,126 @@
----
-license: mit
----
+---
+language:
+- en
+license: mit
+library_name: transformers
+tags:
+- svg
+- vector-graphics
+- text-to-svg
+- image-to-svg
+- hierarchical-tokenization
+- autoregressive-generation
+- code-generation
+base_model: Qwen/Qwen2.5-VL-3B
+pipeline_tag: image-to-text
+datasets:
+- svg-stack
+model-index:
+- name: HiVG-3B-Base
+  results: []
+---
+
+# HiVG-3B-Base
+
+**HiVG-3B-Base** is a 3B-parameter vision-language model for **autoregressive Scalable Vector Graphics (SVG) generation**. It is the base model from the paper [**"Hierarchical SVG Tokenization: Learning Compact Visual Programs for Scalable Vector Graphics Modeling"**](https://arxiv.org/abs/2604.05072).
+
+HiVG introduces a novel **hierarchical SVG tokenization framework** that replaces generic byte-level tokenization with geometry-aware atomic and segment tokens, enabling significantly more efficient and faithful SVG code generation.
+
+| 📄 [Paper](https://arxiv.org/abs/2604.05072) | 🏠 [Project Page](https://hy-hivg.github.io/) | 🤗 [Paper Page](https://huggingface.co/papers/2604.05072) |
+|---|---|---|
+
+## Model Description
+
+### Overview
+
+Recent large language models have shifted SVG generation from differentiable rendering optimization to autoregressive program synthesis. However, existing approaches still rely on **generic byte-level tokenization** inherited from natural language processing, which poorly reflects the geometric structure of vector graphics — numerical coordinates are fragmented into discrete symbols, destroying spatial relationships and inflating token length and computational cost.
+
+**HiVG** addresses these fundamental challenges through a hierarchical SVG tokenization framework:
+
+1. **Atomic Tokens (Level 1):** Raw SVG strings are decomposed into structured atomic tokens that preserve the full geometric semantics of SVG commands (structure, command type, and coordinates).
+2. **Segment Tokens (Level 2):** Executable command–parameter groups are further compressed into geometry-constrained segment tokens, substantially improving sequence efficiency while preserving syntactic validity.
+3. **Hierarchical Mean-Noise Initialization:** A novel embedding initialization strategy that bridges the gap between pre-trained LLM embeddings and the new SVG token space.
+4. **Curriculum Training Paradigm:** A training strategy that progressively increases SVG program complexity, enabling more stable learning of executable SVG programs.
+
+### Architecture
+
+- **Parameters:** ~3B (4B total including vision encoder)
+- **Training Strategy:** Full-parameter Supervised Fine-Tuning (SFT) with **frozen vision encoder**
+- **Tokenization:** Hierarchical SVG tokenizer (atomic + segment tokens)
+
+## Intended Uses
+
+### Primary Use Cases
+
+- **Text-to-SVG Generation:** Generate SVG vector graphics from natural language descriptions.
+- **Image-to-SVG Generation (Vectorization):** Convert raster images into editable SVG code.
+
+### Out-of-Scope Uses
+
+- This is a **base model** and has not been instruction-tuned or RLHF-aligned for production deployment.
+- Not designed for generating arbitrary code beyond SVG.
+- Not suitable for safety-critical applications without additional safeguards.
+
+## Training Details
+
+### Training Procedure
+
+- **Backbone:** Qwen2.5-VL-3B
+- **Fine-tuning:** Full-parameter SFT with frozen vision encoder
+- **Curriculum Learning:** The model was trained with a curriculum training paradigm that progressively increases program complexity
+- **Initialization:** Hierarchical mean-noise initialization strategy for new SVG token embeddings
+
+### Compute Infrastructure
+
+Please refer to the [paper](https://arxiv.org/abs/2604.05072) for detailed compute specifications.
+
+## Evaluation
+
+### Tasks
+
+The model was evaluated on both:
+- **Text-to-SVG** generation
+- **Image-to-SVG** generation (vectorization)
+
+### Results
+
+Extensive experiments demonstrate that HiVG improves:
+- **Generation fidelity** — higher visual quality of rendered SVGs
+- **Spatial consistency** — better preservation of geometric layouts and spatial relationships
+- **Sequence efficiency** — significantly shorter token sequences compared to conventional byte-level tokenization schemes
+
+For detailed quantitative results, tables, and comparisons with baselines (e.g., StarVector, DuetSVG), please refer to the [paper](https://arxiv.org/abs/2604.05072).
+
+## How to Use
+
+```python
+from hivg_infer import HiSVGInferencePipeline
+
+pipeline = HiSVGInferencePipeline(
+    model_path="/path/to/model",
+    coord_range=234,
+    temperature=0.7,
+    top_p=0.9,
+    max_new_tokens=4096,
+)
+
+# Image-to-SVG
+result = pipeline.img2svg("assets/cases/w2.png")
+if result["success"]:
+    print(result["svg"])
+```
+
+> Note: For detailed inference code, data preprocessing, and the hierarchical SVG tokenizer/detokenizer, please visit the [project page](https://hy-hivg.github.io/) and the associated code repository.
+
+## Citation
+
+If you find this work helpful, please cite:
+
+```bibtex
+@article{xing2026hivg,
+  title={Hierarchical SVG Tokenization: Learning Compact Visual Programs for Scalable Vector Graphics Modeling},
+  author={Ximing Xing and Ziteng Xue and Zhenxi Li and Weicong Liang and Linqing Wang and Zhantao Yang and Tiankai Hang and Zijin Yin and Qinglin Lu and Chunyu Wang and Qian Yu},
+  journal={arXiv preprint arXiv:2604.05072},
+  year={2026}
+}
+```