Upload README.md

README.md

@@ -10,16 +10,16 @@ tags:
 - text-to-image
 - svg
 - vector-graphics
+- image-to-vector
+- flow-matching
 ---
 
 # SVG Generator
 
 This repository contains models for a two-stage text-to-SVG generation pipeline.
-The first published component is a LoRA adapter for Z-Image that biases image
-generation toward clean SVG-style illustrations.
-
-The vectorization model will be added as a separate component in the same
-repository.
+It includes a LoRA adapter for Z-Image that biases image generation toward clean
+SVG-style illustrations and a flow-matching vectorizer that converts raster
+images into Bezier-curve SVGs.
 
 ## Z-Image SVG LoRA
 
@@ -77,3 +77,45 @@ SVG illustration with white background.
 This is only the bitmap-generation stage of the full pipeline. The output of
 the LoRA is still a raster image and must be converted to SVG by a separate
 vectorization model or tool.
+
+## Flow-Matching Vectorizer
+
+Vectorizer files:
+
+- `flow-matching/config.json`
+- `flow-matching/model.safetensors`
+
+The vectorizer is conditioned on DINOv3 image features from
+`facebook/dinov3-vits16-pretrain-lvd1689m`. The DINOv3 encoder is not stored in
+this repository; it is loaded separately from its original Hugging Face
+repository.
+
+The model can be loaded from the Hub with the helper code in this repository:
+
+```python
+import torch
+from flow_matching_hf import load_dino_encoder, load_flow_matching_from_hub
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+vectorizer = load_flow_matching_from_hub(
+    "JosefKuchar/svg-generator",
+    subfolder="flow-matching",
+    device=device,
+)
+processor, image_encoder = load_dino_encoder(device=device)
+```
+
+For folder-based PNG to SVG inference:
+
+```bash
+uv run python vectorize_png_folder_model.py ./pngs ./svgs \
+  --model-repo-id JosefKuchar/svg-generator \
+  --batch-size 1 \
+  --steps 50 \
+  --max-segments 256
+```
+
+The exported vectorizer was created from the original PyTorch Lightning
+checkpoint by removing the frozen DINOv3 encoder tensors and trainer state. The
+original checkpoint is not required for inference.