STAR275/tryon-v1-finetuned

🧥 TryOn V1 — Fine-Tuned Virtual Try-On Model (LoRA)

This repository contains TryOn V1, a fine-tuned LoRA-based diffusion model trained for Virtual Try-On generation.
It takes:

• Person Image
• Cloth Image

and generates a realistic try-on output.

---

🔥 Key Features

• Trained using LoRA fine-tuning
• Supports 256×256 resolution
• Works with large try-on datasets
• Handles different clothing types (women wear, men wear, shalwar kameez, etc.)
• Fast inference and lightweight LoRA weights

---

🧪 Training Method

We used LoRA on a diffusion model with the following configuration:

Parameter	Value
Resolution	256 × 256
Batch Size	1 / 2 (A100 optimized)
Optimizer	AdamW
Learning Rate	1e-4
Epochs	Variable (per folder)
Losses	VAE + Perceptual + Style

Training was performed incrementally:

• Dataset split into multiple folders of 200 images each
• Each folder fine-tuned separately
• Weights were updated iteratively while preventing GPU crash
• Best checkpoints were saved after each run

---

📊 Evaluation Metrics

The following metrics were monitored during training:

• SSIM
• LPIPS
• FID
• Training Loss Curve
• Validation Loss Curve

Graphs for these metrics are included in this repository (if pushed from the training notebook).

---