What is Virtual Try-On?

If ID transfer is "subject transfer specialized for people," then swapping outfits is fair to call "subject transfer specialized for clothes."

It's also called Virtual Try-On (VTON).

Especially for product images, consistency like this matters:

  • Patterns and details stay the same
  • The clothes fit the body shape and pose naturally

LoRA

As with anything, the most reliable and flexible approach is to just train a LoRA of the clothes.

Combine it with inpainting and you can change what a specific person is wearing.

catvton-flux

There are plenty of models specialized for VTON tasks (swapping clothes), but here's catvton-flux as a representative example.

The basic idea is the same as IC-LoRA / ACE++: it uses a side-by-side layout.

catvton_flux_lora.json
  • Left side: Person image
  • Right side: Image of clothes you want them to wear + mask

The model generates an "image of the person on the left wearing the clothes on the right" while looking at both.

Instruction-Based Image Editing (Side-by-Side)

Instruction-Based Image Editing models that don't support multi-reference can't, by design, do things like "take an element from image A and put it into image B."

But as with IC-LoRA / ACE++, the side-by-side technique plus a LoRA trained for it gets you something close.

flux_kontext_lora_v0_2.json

The model generates an "image of the person on the left wearing the clothes on the right" while looking at both.

I used my own LoRA as the example because I wanted to show it off, but just one day later a far better LoRA for Qwen-Image-Edit came out ☹️ Clothes Try On (Clothing Transfer) - Qwen Edit

The biggest advantage of using instruction-based image editing is that you don't need a mask.

For example, say you want to put jeans on someone wearing a miniskirt. With regular VTON, you have to mask not just the miniskirt but also the legs that the jeans will cover. Generating a mask for that combined area automatically is really hard.

Instruction-based image editing needs no mask, so you can swap clothes without worrying about any of that hassle.

Instruction-Based Image Editing (Multi-Reference)

With an instruction-based image editing model that supports multi-reference, it's easy.

Just pass the person and the outfit into separate slots and give an instruction like "dress this person in these clothes," and you're done.