Full video tutorial: the free workflow, the BFS V1.1 vs alternative LoRA, and a full sampler x scheduler test: https://youtu.be/km4zxl3gtZ8
Face swap in ComfyUI with the BFS (Best Face Swap) LoRA by Alissonerdx on the official Qwen Image 2.1 Turbo: 8 steps, CFG 1, no speed LoRA. Load a body image and a head reference: it replaces the head with the reference's face and hair and keeps the body, pose, clothes, lighting and background. A built-in side-by-side preview shows body, head and result at a glance.
This gallery: the character from my Katana video, made locally with Qwen Image 2.1 and kept consistent across every shot with this workflow (the Katana video is sponsored by Higgsfield): https://youtu.be/ly6Mj0oFqU0
Settings (already set):
BFS V1.1 head LoRA at strength 1.0 (the V1.1 alternative is in the loader, switched off)
8 steps, CFG 1, sampler lcm, scheduler bong_tangent (from the RES4LYF node pack)
image 1 = body, image 2 = head reference; the swap prompt is already in the text encode node
CFG 1.2 was more than twice as slow in my tests with no gain in resemblance: stay at 1
Timed on an RTX 4000 Ada: about 45 s per swap
Tips
Use a sharp, front-facing head reference, and crop it so no props are in it (an uncropped reference put a second sword on my character's back).
Use a body image where the face is clearly visible (waist-up or closer); faces far away in full-body shots swap less well.
Qwen 2.1 can shift the whole image slightly (colours, softness) even for small edits; higher resolutions help.
Models (Hugging Face):
BFS head LoRAs (V1.1 + alternative) by Alissonerdx: https://huggingface.co/Alissonerdx/BFS-Best-Face-Swap/tree/main
Qwen Image 2.1 Turbo by the Qwen team: https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo
ComfyUI files (Turbo int8, text encoder, VAE) by Comfy-Org: https://huggingface.co/Comfy-Org/Qwen-Image-2.1/tree/main
Low VRAM: Turbo Q5_K_M GGUF by Abiray: https://huggingface.co/Abiray/Qwen-Image-2.1-Turbo-GGUF/tree/main
Custom nodes: ComfyUI-GGUF, rgthree-comfy, ComfyUI-KJNodes (side-by-side preview) and RES4LYF for the bong_tangent scheduler (ComfyUI Manager > Install Missing Custom Nodes): https://github.com/ClownsharkBatwing/RES4LYF
System requirements: NVIDIA RTX GPU, about 8 GB VRAM for the int8 Turbo or 6-8 GB with the GGUF, 20 GB free disk.
Want more realism? The supporters edition adds an automatic face close-up of the head reference, a detail-matching prompt and a realism LoRA, plus a one-click Windows installer and a walkthrough video:
Patreon (enhanced workflow + one-click installer): https://www.patreon.com/TheLocalLab/posts/qwen-image-2-1-172314062
The walkthrough on YouTube: https://youtu.be/UV0DYy3wgDk
Responsible use: only swap faces of people who gave consent. Never use it on public figures or to deceive. Every image here is AI-generated.
Credits: BFS - Best Face Swap LoRA by Alissonerdx (MIT), Qwen Image 2.1 Turbo by the Qwen team, ComfyUI files by Comfy-Org, GGUF by Abiray, RES4LYF by ClownsharkBatwing, KJNodes by kijai, rgthree-comfy by rgthree, ComfyUI-GGUF by city96.
License: Qwen Image 2.1 Turbo uses the Qwen Research License, which allows non-commercial use only. Read it before you use it: https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo
More from me:
Description
Official Qwen Image 2.1 Turbo (8 steps, no speed LoRA), BFS V1.1 head LoRA (the V1.1 alternative is in the loader, switched off), sampler lcm + scheduler bong_tangent, built-in side-by-side preview (body + head + result), GGUF path for low VRAM.