VTON Pipeline

An end-to-end Virtual Try-On pipeline built for Indian fashion e-commerce. Scrapes a product catalogue, classifies garments, generates diverse model images, composites garments onto models, and optionally swaps the model's face with a target identity for personalised try-on.

How it works

Scrape          Classify         Generate Models       Try-On
products   -->  with GPT-4o  --> (body type x      --> composite
from            Vision           skin tone matrix)     garment onto
westside.com                     27 base PNGs          base model
                                                           |
                                                    Face Swap (optional)
                                                    replace model face
                                                    with target identity

Repository Structure

VTON-Pipeline/
|
|-- pipeline/                    Core pipeline scripts (run in order)
|   |-- scrape.py                Step 1: scrape products from westside.com
|   |-- classify.py              Step 2: classify garments with GPT-4o Vision
|   |-- generate_models.py       Step 3: generate 27 base model images
|   |-- tryon_gpt.py             Step 4A: try-on using GPT Image 2 edit
|   `-- tryon_zimage.py          Step 4B: try-on using Z-Image Turbo (faster)
|
|-- face_swap/                   Step 5: identity personalisation
|   |-- swap_replicate.py        Run face swap via Replicate API (easiest)
|   |-- swap_local.py            Run face swap locally with InsightFace
|   |-- swap_hf_space.py         Call a deployed HF Space for face swap
|   |-- run_hf_swap.py           Swap faces on a batch of try-on images
|   |-- run_pipeline.py          Full end-to-end: try-on + face swap
|   |-- models/                  Place inswapper_128.onnx here
|   `-- requirements.txt
|
|-- spaces/                      Hugging Face Space deployments
|   |-- tryon/                   Z-Image Turbo try-on Space
|   |   |-- app.py
|   |   `-- requirements.txt
|   `-- face_swap/               InsightFace face-swap Space
|       |-- app.py
|       |-- requirements.txt
|       `-- packages.txt
|
|-- tests/                       Print quality and fidelity tests
|   |-- print_fidelity.py        GPT Image 2 print accuracy at quality=low
|   `-- sophisticated_print.py   Scenic and ethnic prints at quality=medium
|
|-- utils/                       Optional / advanced tools
|   |-- fix_missing_images.py    Backfill missing product images after scraping
|   `-- tryon_zimage_local.py    Z-Image Turbo local GPU runner (needs 16 GB VRAM)
|
|-- generated_images/            27 pre-generated base model PNGs (Step 3 output)
|-- westside_dataset/            Scraped product JSON and images
|-- prompts.json                 27 model definitions (body type x skin tone)
|-- requirements.txt             Core dependencies
`-- .env.example                 API key template

All scripts are run from the repository root, e.g. python pipeline/scrape.py. Paths inside every script resolve relative to the current working directory.

Quick Start

1. Clone and set up

git clone https://github.com/Ionio-io/VTON-Pipeline.git
cd VTON-Pipeline

python -m venv .venv
source .venv/bin/activate      # Windows: .venv\Scripts\activate

pip install -r requirements.txt

2. Configure API keys

cp .env.example .env
# Edit .env and fill in your keys

Or export directly:

export FAL_KEY=your_fal_key
export OPENROUTER_API_KEY=your_openrouter_key

3. Run the pipeline

python pipeline/scrape.py
python pipeline/classify.py
python pipeline/generate_models.py --resume
python pipeline/tryon_gpt.py

Pipeline Steps

Step 1 — Scrape Products

Scrapes 50 men's and 50 women's products from westside.com using the public Shopify JSON API. Downloads product images and saves metadata.

python pipeline/scrape.py
python pipeline/scrape.py --per-gender 100   # more products
python pipeline/scrape.py --no-images        # metadata only, very fast

Output:

westside_dataset/products_men.json
westside_dataset/products_women.json
westside_dataset/images/{men|women}/{handle}/1.jpg, 2.jpg, ...

Step 2 — Classify Products

Sends each product's images (as base64) to GPT-4o Vision via OpenRouter. Tags each item with category, usability flag, best image index, and reasoning.

export OPENROUTER_API_KEY=your_key
python pipeline/classify.py
python pipeline/classify.py --gender male    # one gender only
python pipeline/classify.py --delay 2        # seconds between API calls

Output:

westside_dataset/classified_men.json
westside_dataset/classified_women.json

If any product images are missing after scraping, run:

python utils/fix_missing_images.py

Step 3 — Generate Base Models

Generates 27 diverse Indian model images covering 5 male and 4 female body types across 3 skin tones, using GPT Image 2 via fal.ai.

export FAL_KEY=your_key
python pipeline/generate_models.py --resume
python pipeline/generate_models.py --gender male   # one gender only

Output: generated_images/{model_id}.png for each of the 27 IDs in prompts.json

Pre-generated images are included in generated_images/ — skip this step if you want to use them directly.

Step 4A — Try-On with GPT Image 2

Uploads the base model PNG and garment reference photos to fal.ai, then calls GPT Image 2 edit to composite the garment onto the model. Best print fidelity; slower and more expensive than Z-Image Turbo.

export FAL_KEY=your_key
python pipeline/tryon_gpt.py

Edit the SAMPLES list at the top of tryon_gpt.py to choose products and models.

Quality settings (set QUALITY in the script):

Setting	Speed	Notes
`low`	~30 s	Good for rapid iteration
`medium`	~2-4 m	Better print reproduction
`high`	~5-8 m	Best quality for delivery

Output: tryon_output/sample{n}_{model_id}_{handle}.png

Step 4B — Try-On with Z-Image Turbo

Faster and cheaper alternative using Z-Image Turbo image-to-image via fal.ai. Garment is described via text prompt — no reference image upload required. Best for high-volume generation where speed matters more than print fidelity.

export FAL_KEY=your_key

# Single try-on
python pipeline/tryon_zimage.py --model M-REC_S3 --product 0 --gender male

# Batch: first 5 men's products across 3 models
python pipeline/tryon_zimage.py --batch --gender male --max-products 5

# Adjust edit strength (lower = model identity better preserved)
python pipeline/tryon_zimage.py --model F-HG_S4 --product 2 --gender female --strength 0.50

Flag	Default	Description
`--model`	M-REC_S3	Model ID from `prompts.json`
`--product`	0	0-based product index in dataset JSON
`--gender`	male	`male`, `female`, or `both`
`--strength`	0.55	Edit strength (lower = more original kept)
`--batch`	off	Run multiple products x models
`--max-products`	5	Max products per gender in batch mode

Output: tryon_output/zimage_fal/{model_id}_{handle}.png

Cost: ~$0.004 per 768x1024 image on fal.ai

For local GPU inference (16 GB VRAM required), see utils/tryon_zimage_local.py.

Step 5 — Face Swap

Replaces the model's face with a target identity for personalised try-on. Three options depending on your setup:

Option A: Replicate API (recommended, no local GPU needed)

export REPLICATE_API_TOKEN=your_token
python face_swap/swap_replicate.py \
  --source generated_images/M-REC_S3.png \
  --target tryon_output/sample1_M-REC_S3_*.png

Option B: Hugging Face Space

Deploy spaces/face_swap/ as a Gradio Space, then call it:

pip install gradio_client
python face_swap/swap_hf_space.py \
  --space YOUR_USERNAME/YOUR_SPACE \
  --source generated_images/M-REC_S3.png \
  --target tryon_output/sample1_M-REC_S3_*.png

Option C: Local InsightFace (Python 3.10 or 3.11)

pip install -r face_swap/requirements.txt
# Place inswapper_128.onnx in face_swap/models/
python face_swap/swap_local.py \
  --model face_swap/models/inswapper_128.onnx \
  --source generated_images/M-REC_S3.png \
  --target tryon_output/sample1_M-REC_S3_*.png

Full end-to-end pipeline

python face_swap/run_pipeline.py \
  --face your_photo.jpg \
  --product 0 \
  --gender male \
  --model M-REC_S3

HF Space Deployments

Try-On Space (`spaces/tryon/`)

Gradio app that runs Z-Image Turbo image-to-image inference. Upload a base model image, describe the garment, get a try-on result.

Deploy:

Create a new Space at huggingface.co/new-space
- SDK: Gradio
- Hardware: T4-small (16 GB, ~$0.40/hr) or A10G
Upload spaces/tryon/app.py and spaces/tryon/requirements.txt

API:

from gradio_client import Client

client = Client("YOUR_USERNAME/YOUR_SPACE")
result = client.predict(
    base_model_image,
    "Navy Blue Slim-Fit Shirt",
    "Cotton, button-down collar",
    "man",
    "Rectangle",
    "Medium / Olive",
    0.55,   # strength
    42,     # seed
    api_name="/predict",
)

Face Swap Space (`spaces/face_swap/`)

Gradio app powered by InsightFace inswapper_128. Upload source (face to take) and target (face to replace), get swapped result.

Deploy:

Create a new Space — SDK: Gradio, Hardware: CPU Basic
Upload spaces/face_swap/app.py, requirements.txt, packages.txt

Model Matrix

27 base model images covering Indian body types and skin tones:

ID	Gender	Body Type	Skin Tone
M-REC_S3	Male	Rectangle	Medium / Olive
M-REC_S4	Male	Rectangle	Light Brown
M-REC_S5	Male	Rectangle	Dark Brown
M-INV_S3	Male	Inverted Trapezoid	Medium / Olive
M-INV_S4	Male	Inverted Trapezoid	Light Brown
M-INV_S5	Male	Inverted Trapezoid	Dark Brown
M-TRI_S3	Male	Triangle	Medium / Olive
M-TRI_S4	Male	Triangle	Light Brown
M-TRI_S5	Male	Triangle	Dark Brown
M-OVL_S3	Male	Oval	Medium / Olive
M-OVL_S4	Male	Oval	Light Brown
M-OVL_S5	Male	Oval	Dark Brown
M-TRP_S3	Male	Trapezoid	Medium / Olive
M-TRP_S4	Male	Trapezoid	Light Brown
M-TRP_S5	Male	Trapezoid	Dark Brown
F-HG_S3	Female	Hourglass	Medium / Olive
F-HG_S4	Female	Hourglass	Light Brown
F-HG_S5	Female	Hourglass	Dark Brown
F-REC_S3	Female	Rectangle	Medium / Olive
F-REC_S4	Female	Rectangle	Light Brown
F-REC_S5	Female	Rectangle	Dark Brown
F-SPO_S3	Female	Sporty	Medium / Olive
F-SPO_S4	Female	Sporty	Light Brown
F-SPO_S5	Female	Sporty	Dark Brown
F-TRAP_S3	Female	Trapezoid	Medium / Olive
F-TRAP_S4	Female	Trapezoid	Light Brown
F-TRAP_S5	Female	Trapezoid	Dark Brown

API Keys

Script	Service	Environment Variable	Get Key
`pipeline/generate_models.py`	fal.ai	`FAL_KEY`	fal.ai/dashboard/keys
`pipeline/tryon_gpt.py`	fal.ai	`FAL_KEY`	fal.ai/dashboard/keys
`pipeline/tryon_zimage.py`	fal.ai	`FAL_KEY`	fal.ai/dashboard/keys
`pipeline/classify.py`	OpenRouter	`OPENROUTER_API_KEY`	openrouter.ai/settings/keys
`face_swap/swap_replicate.py`	Replicate	`REPLICATE_API_TOKEN`	replicate.com/account/api-tokens
`face_swap/swap_hf_space.py`	Hugging Face (private spaces only)	`HF_TOKEN`	huggingface.co/settings/tokens

Technology Stack

Component	Technology
Base model generation	GPT Image 2 via fal.ai
Try-on (GPT)	GPT Image 2 Edit via fal.ai
Try-on (Z-Image)	Z-Image Turbo via fal.ai
Product classification	GPT-4o Vision via OpenRouter
Face swap (cloud)	InsightFace inswapper_128 via Replicate
Face swap (local)	InsightFace + ONNX Runtime
Data scraping	Westside Shopify public JSON API

Navigation Menu

Uh oh!

Repository files navigation

VTON Pipeline

How it works

Repository Structure

Quick Start

1. Clone and set up

2. Configure API keys

3. Run the pipeline

Pipeline Steps

Step 1 — Scrape Products

Step 2 — Classify Products

Step 3 — Generate Base Models

Step 4A — Try-On with GPT Image 2

Step 4B — Try-On with Z-Image Turbo

Step 5 — Face Swap

Option A: Replicate API (recommended, no local GPU needed)

Option B: Hugging Face Space

Option C: Local InsightFace (Python 3.10 or 3.11)

Full end-to-end pipeline

HF Space Deployments

Try-On Space (`spaces/tryon/`)

Face Swap Space (`spaces/face_swap/`)

Model Matrix

API Keys

Technology Stack

About

Releases

Packages

Used by

Contributors

Languages

Name		Name	Last commit message	Last commit date
Latest commit History 4 Commits
face_swap		face_swap
generated_images		generated_images
pipeline		pipeline
spaces		spaces
tests		tests
utils		utils
westside_dataset		westside_dataset
.env.example		.env.example
.gitignore		.gitignore
README.md		README.md
prompts.json		prompts.json
requirements.txt		requirements.txt

Uh oh!

Folders and files

Latest commit

History

Repository files navigation

VTON Pipeline

How it works

Repository Structure

Quick Start

1. Clone and set up

2. Configure API keys

3. Run the pipeline

Pipeline Steps

Step 1 — Scrape Products

Step 2 — Classify Products

Step 3 — Generate Base Models

Step 4A — Try-On with GPT Image 2

Step 4B — Try-On with Z-Image Turbo

Step 5 — Face Swap

Option A: Replicate API (recommended, no local GPU needed)

Option B: Hugging Face Space

Option C: Local InsightFace (Python 3.10 or 3.11)

Full end-to-end pipeline

HF Space Deployments

Try-On Space (spaces/tryon/)

Face Swap Space (spaces/face_swap/)

Model Matrix

API Keys

Technology Stack

About

Resources

Stars

Watchers

Forks

Releases

Packages

Used by

Contributors

Languages

Try-On Space (`spaces/tryon/`)

Face Swap Space (`spaces/face_swap/`)