Put your character into another photo, or swap only the head. It also works when other people are in the photo: say who to replace, and everyone else stays as they are.

The swaps run on the BFS LoRAs by Alissonerdx (body swap and head swap). The optional Speed LoRA is PrunaAI's 8-step LoRA, converted for ComfyUI by t8star.

How to use it

1. Load your scene, the photo you want to change.

2. Load your character, the person you want in it.

3. In Who to replace, say who gets swapped. With one person in the photo, leave "the person". With more, say where they stand, like "the woman on the left".

4. Run. The request sent to Qwen shows in Show Text, above the result.

To swap only the head, switch Swap the whole body off. The scene's clothes stay.

To swap two people at once, select the Second character nodes, press Ctrl+B, load the second picture and say who it replaces.

Faster, optional

The Speed LoRA is switched off. Download the file below, switch on its first row in both swap LoRA Loaders (Body and Head), and set Steps to 8. A run drops from about 22 seconds to 8 on an RTX 5090, and in my tests the swap looks the same. Only use 8 steps with the Speed LoRA on. Without it, 8 steps looks soft. Leave Steps at 25 if the row is off.

What you need

• AusBoss nodes 2.5.1 or newer. In ComfyUI-Manager, search "AusBoss". Already have them? Update first.

• A ComfyUI with Text Encode Qwen Image 2.1. I tested on 0.38.0.

• The Qwen Image 2.1 models and the two LoRAs:

  • qwen_image_2.1_int8_convrot.safetensors → models/diffusion_models · 7.26 GB

  • qwen3vl_8b_int8_convrot.safetensors → models/text_encoders · 9.35 GB

  • qwen_image_2.1_vae_bf16.safetensors → models/vae · 676 MB

  • bfs_body_swap_v1.0_qwen_2.1.safetensors → models/loras · 210 MB

  • bfs_head_v1.1_qwen_2.1.safetensors → models/loras · 260 MB

Both LoRA files have to be installed, even if you only use one. ComfyUI checks them when you run.

Optional, for the Speed LoRA: p_qwen_image_2.1_8step_v0.1-comfyui-T8.safetensors → models/loras · 336 MB. The row is off, so the workflow runs without the file.

Tips

• Say who by where they stand or by their hair, not by their clothes. For a whole-body swap, clothes you name are kept. "The woman on the right with the braid and yellow top" kept the yellow top. "The woman on the right" swapped it.

• Use a character picture that shows the outfit from head to knees, with the hands down. If it stops at the chest, the old legs and shoes can stay. A hand on the face or a peace sign can move into your scene and add an extra arm. A plain background works best.

• Keep signs and cups out of the character picture. They can follow the person into the scene.

• Whole body takes the character's clothes and build. The scene's pose, light and background stay. Head only keeps the scene's clothes, and on wide shots with small people it barely shows.

• If one person comes out wrong, change the seed. Quiet scenes with one or two people sitting or standing work best.

• Only use pictures of people who agreed to it. The LoRA's author asks for this.

• About 20 seconds a run for one character, 28 for two, on an RTX 5090. With the Speed LoRA, about 8.

If a swap goes wrong

The Add to the request node is switched off. It holds three sentences that go on the end of the request. Select it and press Ctrl+B to switch it on, and delete any sentence you don't need.

• The scene's old legs and shoes stay (the character picture stops at the chest): "replace the legs and shoes as well, nothing of the old outfit from <image1> stays."

• A hand from the character picture shows up (a hand on the face, a peace sign): "keep the arms and hands exactly as in <image1>."

• A sign, cup or other object from the character picture comes along: "do not copy signs, poles, cups or other objects from <image2>."

<image1> is your scene and <image2> is your character. The hands sentence did not fix small people in a wide shot.

Credits

Qwen Image 2.1 is by Alibaba Qwen, under the Qwen Research License: non-commercial use only. The BFS LoRAs are by Alissonerdx, MIT License. The optional 8-step Speed LoRA is by PrunaAI, converted for ComfyUI by t8star, under the same Qwen Research License: research use only.

v1.0 (2026-10-02): first release.

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