A consistency LoRA for outfit swaps with Qwen Image 2.1: the person gets the outfit from a second picture, and the rest of the picture stays where it was.

Qwen Image 2.1 can already dress a person in the outfit from another picture. The result comes back a few pixels off, the face a little softer, the background repainted. With this LoRA only the clothes change.

No trigger word. Load it at strength 1.0, give the person as image 1 and the outfit as image 2, and ask for the swap:

Dress the person in image 1 in the <outfit, in a few words> shown in image 2. Keep their face, hair, hands, pose and the background exactly the same.

Why use it

Qwen Image 2.1 can do an outfit swap without this LoRA. This is what the LoRA changes, measured on 16 swaps it never trained on, same seed and request:

• The picture stays in place. Without the LoRA the result sits 3.6 px away from your picture, so it no longer lines up with the original. With it: 0.1 px.

• The face stays the same. Without: 14.1 % of the face's pixels change noticeably. With: 1.7 %.

• The background is left alone. Without: 10.8 % of it gets repainted. With: 0.9 %.

• The old outfit comes off. Without, old boots or jeans stay on under the new outfit (cleared in 0 of 3 seeds). With: 2 of 3.

The pictures on this page show it. Yellow marks every pixel that differs from the original picture. Without the LoRA the face, the hair and the background light up. With it, only the clothes do.

What happens to the old clothes

A sentence after the request steers what stays on:

• Replace everything they wear. takes off old tights and legwear.

• Keep their own shoes and legwear. keeps them.

How to use it

My Outfit Swap workflow has it set up: two pictures in, the outfit gets described for you, the sentence above sits in a box you can edit.

In your own two-picture Qwen Image 2.1 edit graph:

1. A LoRA loader right after the model loader, strength 1.0. At 0.5 the picture starts to move again.

2. Text Encode Qwen Image 2.1: the person as image_1, the outfit as image_2, resolution 0, the request as the prompt.

3. KSampler on the encoder's latent output: 25 steps, CFG 1, euler / simple, denoise 1.

4. VAE Decode, then Split Image with Alpha (the Qwen 2.1 VAE decodes RGBA).

Limits

• It holds the picture, so it holds the pose. An outfit that needs a different pose will not get one.

• Shoes are the stubborn part. A pair sometimes stays on either way.

• Something that lies over the clothes, like a bag strap or an earphone cable, can go with the old outfit or sit oddly on the new one.

• Trained at about 1 MP, tested at about 2 MP with 25 steps and CFG 1.

• With the Viggle Turbo LoRA at 8 steps it still holds the picture. Plain fabric can come out a little grainy.

Training

• ai-toolkit (qwen_image_2) on the Comfy-Org INT8 convrot base. Rank 32, alpha 32, AdamW8bit, learning rate 1e-4, batch 1, 1,000 steps.

• 66 swap pairs. Each started as a real Qwen Image 2.1 swap. I moved it back onto the person picture by whole pixels and kept only the clothes from the swap: the face, hair, hands and background are the person picture's own pixels. I looked at every one and kept the ones I liked.

The full numbers, the step 1,250 and 1,500 files and the notes on how it was made are on the Hugging Face model card.

Qwen Image 2.1 is by Alibaba Qwen, under the Qwen Research License: non-commercial use only.

Questions or bugs: the comments here or GitHub issues. More from me on X @Zanzibased and GitHub.