Skip to main content

TryOnDiffusion Overview

TryOnDiffusion is a diffusion-based virtual try-on model using a dual UNet architecture.

Key Features​

  • Dual UNet architecture (Person UNet + Garment UNet)
  • Pose-conditioned generation
  • Cross-attention between person and garment features
  • Support for 64x64 and 128x128 resolutions

Architecture​

The model consists of two parallel UNets:

  1. Person UNet: Generates the final output image
  2. Garment UNet: Processes segmented garment information

Quick Start​

from tryondiffusion.diffusion import Diffusion

diffusion = Diffusion(
    device="cuda",
    pose_embed_dim=8,
    time_steps=256,
    unet_dim=64
)

See Training Guide and Inference Guide for details.