How to Install BFS Face Swap with Qwen Image 2.1 in ComfyUI
Install BFS, a face-swap LoRA for Qwen Image 2.1, in ComfyUI. Covers model downloads, VRAM use, custom nodes and workflow setup.
What is BFS and why is it getting attention?
BFS, short for Best Face Swap, is a collection of LoRA files built for head, face, and body replacement across several base image and video models. It’s published on Hugging Face by the creator Alissonerdx under the repo “BFS-Best-Face-Swap,” and as of this writing it has racked up over 212,000 downloads and nearly 1,200 likes, which puts it among the more popular face-swap tools circulating in the open-source ComfyUI community.
Rather than being a single monolithic model, BFS ships as a set of LoRA adapters trained for different backbones: Qwen Image 2.1, Qwen Image Edit 2509, Flux 2 Klein (4B and 9B variants), Krea 2, and even LTX-2 for video-based head swaps. A LoRA (low-rank adaptation) is a small add-on file that nudges a large base model toward a specific task without retraining the whole thing. In this case, the LoRA teaches the base model how to graft one person’s face or head onto another image while preserving pose, lighting, and background.
How does the Qwen Image 2.1 head-swap workflow work?
The version most people are installing right now pairs BFS with Qwen Image 2.1, Alibaba’s image generation and editing model that’s become a staple in local AI art workflows for its editing capabilities and open license.
One coffee. One working app.
You bring the idea. Remy manages the project.
The basic pipeline works like this: you feed the workflow a reference image (the scene or pose you want to keep) and a source face image (the face you want swapped in). The Qwen Image 2.1 base model handles the overall image generation and editing, while the BFS LoRA constrains it to perform a head swap, keeping the original lighting, camera angle, and surrounding context intact. The result is a composite that looks like the subject was actually in the original photo, not pasted on top of it.
In practical testing, output quality depends heavily on the input face image. Reference photos with good lighting and a direct, forward-facing angle produce noticeably cleaner, more realistic results than low-quality or off-angle source images. This is consistent with how diffusion-based face and head swap models generally behave: garbage in, garbage out still applies even with a well-trained LoRA doing the heavy lifting.
What do you need to install before downloading BFS?
Before touching the BFS files, you need a working ComfyUI installation with the ComfyUI Manager extension installed, since Manager is used to pull in the custom nodes BFS depends on.
You also need the Qwen Image 2.1 base model stack, which typically consists of three separate files:
- A diffusion model file (the core generative weights)
- A VAE (variational autoencoder, which handles encoding and decoding images to and from latent space)
- A text encoder (which processes your text prompts)
These get downloaded individually from the Qwen Image 2.1 model repository and placed into their respective ComfyUI subdirectories: diffusion models go in the diffusion models folder, the VAE goes in the VAE folder, and the text encoder goes in the text encoders (or clip) folder. If you’ve never set up Qwen Image 2.1 before, it’s worth working through a dedicated Qwen Image 2.1 installation guide first, since getting the folder structure wrong is the most common source of “model not found” errors in ComfyUI.
How do you install the BFS LoRA files themselves?
Once the Qwen Image 2.1 base model is in place, installing BFS is comparatively simple:
- Go to the BFS-Best-Face-Swap repository on Hugging Face and open the “Files” tab.
- For the Qwen Image 2.1 head-swap version, download the corresponding
.safetensorsfile (the repo includes multiple versions, such asbfs_head_v1_qwen_2.1.safetensorsand newer iterations like v2 through v5, plus a merged 2511 version in different rank and precision options). - Right-click the download arrow, choose “Save link as,” and save the file into your ComfyUI
models/lorasfolder. - Optionally, there’s a “pruned” (lighter) variant of the LoRA available if you want a smaller file footprint or faster loading.
After the LoRA is in place, open ComfyUI’s Manager, go to Custom Nodes, and install the node package associated with BFS (the installation guide specifically calls out needing the “address for life” custom node, referring to the specific custom node pack the BFS workflow JSON relies on). Installing it through Manager and restarting ComfyUI completes the setup.
From there, load the official BFS workflow JSON (available on the model card alongside the files), point the loader nodes at your downloaded models, drop in your reference image and source face image, and run the generation.
How much VRAM does BFS with Qwen Image 2.1 use?
Running the full Qwen Image 2.1 plus BFS head-swap pipeline end to end consumes just under 23GB of VRAM. That places it squarely in high-end consumer or prosumer GPU territory (think 24GB cards), and it rules out most mainstream 8GB to 12GB cards without further quantization or offloading tricks.
Remy doesn't write the code. It manages the agents who do.
Remy runs the project. The specialists do the work. You work with the PM, not the implementers.
If you’re on a lower-VRAM setup, look at the pruned LoRA variant and consider using quantized versions of the Qwen Image 2.1 diffusion model, though that tradeoff will cost you some generation quality or speed. Generation time itself was reported as fast once the pipeline is loaded, with individual swaps completing in roughly 20 seconds on hardware capable of handling the 23GB load.
Is BFS worth setting up locally?
For anyone already running Qwen Image 2.1 locally in ComfyUI, BFS is a relatively low-effort add-on that produces results noticeably better than many older face-swap LoRAs, particularly around preserving lighting, pose, and background consistency. Multiple test swaps across different reference scenes (climbing photos, portraits, group-style shots) showed the model correctly retaining the scene’s character while replacing the head, including subtler details like matching skin tone to ambient lighting.
That said, it’s not flawless. Occasional artifacts showed up in testing, such as unexpected changes to facial expression or small distortions around the mouth that weren’t present in either source image. Output quality is also clearly tied to the quality of your input face photo: a poorly lit or off-angle reference image produces a visibly worse swap than a clean, front-facing one.
The real cost is hardware. Needing close to 23GB of VRAM means this isn’t a casual download for most laptop or mid-range desktop GPU users. If you’ve got the VRAM headroom and already have Qwen Image 2.1 set up, BFS is a worthwhile LoRA to add to your toolkit. If you don’t, the base model requirement alone is the bigger barrier.
What are the ethical considerations with face-swap tools like this?
Face-swap models carry obvious misuse risk, and that risk doesn’t go away just because the tool runs locally and open source. Responsible use means only swapping in your own face or faces of people who’ve explicitly consented, never using the tool on celebrities, public figures, or private individuals without permission, and never using it to impersonate, deceive, or embarrass someone. Content generated with tools like BFS should be clearly labeled as AI-generated or synthetic when shared publicly, in line with platform policies like YouTube’s altered/synthetic content disclosure requirements.
Frequently Asked Questions
What is BFS in the context of AI face swapping?
BFS (Best Face Swap) is a set of LoRA adapter files hosted on Hugging Face that add head, face, and body swap capability to several base image and video generation models, including Qwen Image 2.1, Qwen Image Edit, Flux 2 Klein, Krea 2, and LTX-2.
Do I need to train my own model to use BFS?
No. BFS is a pre-trained LoRA, meaning you download a single small file and load it on top of an already-installed base model like Qwen Image 2.1. No training or fine-tuning is required on your end.
How much VRAM does BFS with Qwen Image 2.1 require?
Running the full pipeline, base model plus BFS LoRA, end to end consumes just under 23GB of VRAM, which generally requires a 24GB-class GPU.
Can BFS swap faces in video, not just images?
Yes, the BFS repository includes a head-swap LoRA variant trained for LTX-2, a video model, in addition to the image-focused versions built for Qwen Image 2.1, Qwen Image Edit, Flux 2 Klein, and Krea 2.
Is it legal or safe to use face-swap tools like BFS?
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
The tool itself is legal to download and run, but using it on anyone’s likeness without their clear consent, especially public figures or to deceive or embarrass someone, raises serious ethical and potentially legal issues. Best practice is to use your own face, get explicit consent from others, and label any shared output as AI-generated.