Open source · v0.0.7
Open-source AI for fashion and virtual try-on.
One model registry, four ways to run it. Drive 148 models across 10 services from the CLI, an MCP server, TryOn Studio or Python.
- models
- 148
- services
- 10
- ways to run
- 4
CLI
Run any registry model from the shell with one consistent flag set and --dry-run.
opentryon vton --model fashn-vton-1.5MCP server
Every registry model is exposed as a tool for Claude, Cursor and other MCP clients.
python mcp-server/server.pyTryOn Studio
An agent chat and capability screens that talk to the MCP server over HTTP.
Studio → MCP → modelsPython APIs
Adapters under tryon.api and a single invoke_model() entry point.
from tryon.cli.runner import invoke_modelWelcome to OpenTryOn
OpenTryOn is an open-source AI toolkit for fashion technology and virtual try-on. Current release: v0.0.9 on PyPI.
🎯 What is OpenTryOn?
OpenTryOn gives you four ways to run fashion AI models:
- CLI —
opentryon <service> --model <model> … - MCP server — tools for Cursor, Claude, and any MCP client (guide, in-repo README)
- TryOn Studio — Next.js playground that talks to the MCP server (setup and screens)
- Python APIs —
tryon.apiadapters +invoke_model()
Plus preprocessing, datasets, Gradio demos, and TryOnDiffusion research code.
What's new in v0.0.9
- Decision models: Perplexity Decider, Cloudflare Clef, Kev, AutoTrust JEV, vLLM-SR D3 and Mercury Decide on one
decidequestion schema (guide) - New services:
robot-act(Dream-1, Cosmos 3 Nano-DROID),detect(Moondream boxes, points, masks) andparse(Nemotron Parse) - Gemini 4 Argon, Kimi K3.1 Pro, Moondream 3.1 and Mercury 3 on
understand - Gemini Omni 1.2 (now the default, up to 15s), FLUX 3 Max, Kling 4.0, Seedream 5.0 Flash, and Nano Banana video-to-image
What was new in v0.0.8
measure— Clad body sizing from a photo, a questionnaire, or local Anny/MHR params- FLUX 3 Image Klein (API + local) and FLUX 3 Video
audio-to-video(nvidia-lipsync) andgenerate-3d(trellis)- Nemotron 3.5 Super VL, VideoJev (local video decisions), Gemini Embedding 2
🚀 Key Features
Developer surfaces
- Unified registry-driven CLI with
--dry-run - FastMCP server — every registry model is a tool (MCP Server)
- TryOn Studio — Agent chat, Connect, and capability screens over MCP HTTP (setup)
- OpenAPI / Swagger + Postman snapshots for upstream media APIs (guide)
- Model integration guidelines for Path A (API) vs Path B (local)
Virtual try-on
Cloud adapters including FLUX VTO, Google Vertex Virtual Try-On, OutfitAnyone-Plus, Photoroom (try-on + virtual model), Nova Canvas, Kling AI, Segmind, Pruna P-Image-Try-On, FASHN, Nano Banana 2 Lite composition, Qwen-Image (API + local), Leffa and CatVTON (local weights), and Muse Image composition.
Image generate / edit
Nano Banana family, FLUX.2, GPT Image (1.5 + ChatGPT Images 2.5 Flare/Sunburst), Luma Photon, Seedream 5.0 Pro, Ideogram 4.0, Grok Imagine Image, Pruna P-Image / P-Image-Ideogram / Edit / Upscale, Qwen-Image (DashScope 3.0 + local 2512/Edit-2511), Muse Image, plus local FLUX.2-dev Turbo.
Video
Veo, Sora, Luma Ray 2 + Ray 3.2, Seedance 2.5, Kling 3.0 / Omni / Turbo, Grok Imagine Video, Gemini Omni 1.2 Flash, Pruna P-Video / Replace / Avatar / Animate, plus LTX-2.5 (API + local), Hailuo 2.3, MiniMax H3 / H3 Max (API + local H3 + Fal H3 Max + Relight), Wan (2.x API + Wan 3.0 API + local 2.2), Runway Gen-4.5.
Understanding & other
Claude Haiku 5.5 (API), Kimi K2.6 / K2.7 Code / K3 (API), Kimi-VL & LLaVA-NeXT (local), Qwen3.8-Max (API) + Qwen3.8-27B (local), NVIDIA Nemotron Omni / Cosmos 3 Reasoner, Hy4 preview (TokenHub + local vLLM/SGLang), BEN2 bg-remove, fashion datasets, garment/pose preprocessing. Also NVIDIA Cosmos 3 video generation.
Decisions & embeddings
New decide service (calibrated yes/no, choice, score answers): Typesafe Jev (API), Liquid AI d1-3B / d1-omni-600M (local). New embed service: EmbeddingGemma 2 and pplx-embed-v2-late 0.6B / 9B (local; vectors saved to .npz).
Interactive playground
TryOn Studio is the Next.js UI: Agent chat (planner_agent), Connect (MCP status + key passthrough), and capability screens (Image, VTON, Understand, Video, BG Remove). In-repo Gradio apps remain for extract-garment / model-swap / outfit-generator.
📚 What You'll Learn
In this documentation, you'll find:
- Installation Guide: Get OpenTryOn up and running (
pip install opentryon) - Quick Start: First successful runs
- Configuration: API keys and
.env - Unified CLI: Service → model → params
- MCP Server: Agent / IDE tool surface, plus the
mcp-serverREADME - TryOn Studio: Web UI setup, Agent, Connect, and capability screens
- OpenAPI & Postman: Swagger for upstream media APIs
- Fashion ML Engineer Path: Train → eval → invoke → workflow (toward v0.1.0)
- Datasets Module: Fashion-MNIST, VITON-HD, Subjects200K
- API Reference: Adapters and provider docs
- Examples: Real-world usage examples
- Advanced Guides: Training and customization
- Roadmap: Shipped vs remaining work
🎓 Prerequisites
Before you begin, you should have:
- Python 3.10 or higher
- Basic knowledge of Python programming
- Familiarity with computer vision concepts (helpful but not required)
- CUDA-capable GPU (recommended for best performance)
💡 Quick Examples
Here are some simple examples to get you started:
Virtual Try-On with Segmind
from dotenv import load_dotenv
load_dotenv()
from tryon.api import SegmindVTONAdapter
# Initialize adapter
adapter = SegmindVTONAdapter()
# Generate virtual try-on
images = adapter.generate_and_decode(
model_image="person.jpg",
cloth_image="shirt.jpg",
category="Upper body"
)
# Save result
images[0].save("result.png")
Using Fashion-MNIST Dataset
from tryon.datasets import FashionMNIST
# Create dataset instance (downloads automatically)
dataset = FashionMNIST(download=True)
# Load the dataset
(train_images, train_labels), (test_images, test_labels) = dataset.load(
normalize=True,
flatten=False
)
print(f"Training set: {train_images.shape}") # (60000, 28, 28)
print(f"Class 0: {dataset.get_class_name(0)}") # 'T-shirt/top'
Garment Preprocessing
from dotenv import load_dotenv
load_dotenv()
from tryon.preprocessing import segment_garment, extract_garment
# Segment garment
segment_garment(
inputs_dir="data/original_cloth",
outputs_dir="data/garment_segmented",
cls="upper"
)
# Extract garment
extract_garment(
inputs_dir="data/original_cloth",
outputs_dir="data/cloth",
cls="upper",
resize_to_width=400
)
🤝 Get Involved
OpenTryOn is an open-source project, and we welcome contributions!
- GitHub: github.com/tryonlabs/opentryon
- Discord: Join our community
- Contributing: See our Contributing Guide
📄 License
All material is made available under Creative Commons BY-NC 4.0. You can use the material for non-commercial purposes, as long as you give appropriate credit and indicate any changes.
🗺️ Roadmap
Check out our Roadmap to see what's coming next!
🆘 Need Help?
- Check our Troubleshooting Guide
- Join our Discord community
- Open an issue on GitHub
Ready to get started? Head over to the Installation Guide!