crossimage-tryon-fluxkontext

Cross-Image Try-On Flux Kontext

Model description

Cross-Image Try-On LoRA (Flux Kontext)

This LoRA enables experimental virtual try-on by leveraging cross-image context with Flux Kontext.
By placing a subject and a reference outfit side-by-side in a single image, the model attempts to transfer the clothing style from one to the other.

⚠️ This model is for research/experimental purposes only. It lacks the stability and accuracy required for real-world applications.

Usage

• Prepare an image with the reference outfit (left) and the target person (right).
• Use one of the following trigger prompts:
  • Full-body:
    Change all clothes on the right to match the left
  • Upper-body only:
    Change the upper body clothing on the right to match the left
• Adding explicit details in the prompt can improve reliability, e.g.
  "right woman wearing a red dress".

ComfyUI workflow example is included: Flux Kontext LoRA.json

Dataset (v0.2)

• 53 paired images
• Subject images from Chroma1-HD
• Outfit references from catvton-flux
• Each pair: reference outfit (left) + target person (right, edited with the left outfit)
• Example pairs and workflow in the training/ folder

Training (v0.2)

• Trainer: AI-ToolKit
• Learning rate: 5e-5
• ~65,000 steps (likely fewer would suffice)

Limitations

• Clothing transfer may fail depending on outfit or context
• Works best on realistic photos; illustration/anime styles generally fail
• Success rate improves if the prompt explicitly describes the clothing, e.g.
  "right woman wearing a white shirt"
• Dataset size is reasonable, but more variety (outfits, poses, styles) would likely help

Version History

• v0.2 (current)

  • Expanded dataset (53 pairs)
  • Sources: Chroma-HD (subjects), catvton-flux (outfits)
  • Trained with AI-ToolKit (LR 5e-5, ~65k steps)
  • Added separate full-body & upper-body trigger phrases

• v0.1

  • Small dataset from Pexels
  • Trained with fal-ai/flux-kontext-trainer (default settings)
  • Single trigger phrase:
    Change the clothes on the right to match the left