Open-Source Face and Head Swap LoRA 'BFS-Best-Face-Swap' Adds Qwen Image 2.1 Support

An in-depth look at BFS-Best-Face-Swap, an open-source LoRA suite leveraging Qwen-Image-2.1's multimodal DiT architecture to achieve consistent face, head, and

tau · October 3, 2026

#qwen-image-2.1 #face-swap #lora #comfyui #diffusers #dev-tools

Open-Source Face and Head Swap LoRA 'BFS-Best-Face-Swap' Adds Qwen Image 2.1 Support

Open-source AI developer Alissonerdx has released an update to the 'BFS-Best-Face-Swap' LoRA collection on Hugging Face, officially introducing support for Alibaba's 7B diffusion transformer model, Qwen-Image-2.1. By combining the native multimodal context capabilities of Qwen-Image-2.1 with specialized fine-tuning, the project addresses longstanding issues in local portrait editing pipelines, such as boundary discoloration, unnatural skin smoothing, and lighting mismatches.

Conceptual diagram of BFS-Best-Face-Swap LoRA pipeline on Qwen-Image-2.1 with ComfyUI node interface and character replacement stages

Image source: Alissonerdx / Hugging Face

The BFS (Best Face Swap) repository extends beyond conventional single-layer face swapping, offering granular replacements across individual facial features, full head transfers, and body swaps. The project packages ready-to-run ComfyUI workflow JSON files, Draw Things presets, and Python Diffusers deployment documentation for local reproducibility.

Three-Tier Replacement Architecture: Face, Head, and Body

BFS-Best-Face-Swap structures character editing into three distinct operational modes:

  • Face Swap: Swaps central facial features (eyes, nose, and mouth) while strictly preserving the original image's hair, underlying skin tone, and micro-expressions.
  • Head Swap: Replaces the entire head, including hairstyles and accessories, seamlessly blending the neckline into the target subject's body geometry and posture.
  • Body Swap: Transfers the reference subject's body into a target scene while preserving original background composition, perspective, and ambient illumination.

Because Qwen-Image-2.1 natively processes multi-image conditioning (supporting up to 10 reference images in a single context), the BFS LoRA leverages this structural spatial awareness to resolve 3D head rotation and directional lighting. This approach eliminates the harsh seam lines and inverted shadows frequently observed in older Stable Diffusion-based inpainting setups.

ComfyUI Workflow, Prompt Token Formatting, and Hardware Profile

The repository provides designated weights such as bfs_head_v1.1_qwen_2.1.safetensors alongside a turnkey ComfyUI workflow.

Effective execution requires strict adherence to reference token notation (<image1> and <image2>) within the generation prompt:

head_swap: start with <image1> as the base image, keeping its lighting, environment, and background.
remove the head from <image1> completely and replace it with the head from <image2>,
strictly preserving the hair, eye color, nose structure from <image2>.
copy the direction of the eye, head rotation, micro expressions from <image1>,
high quality, sharp details, 4k
  • Hardware Requirements: Running the workflow locally targets approximately 10GB to 12GB of VRAM (NVIDIA GeForce RTX 4070 or equivalent recommended).
  • Inference Parameters: Recommended settings include the Euler sampler with Simple scheduling, 12 generation steps, and a CFG Scale of approximately 2.5.

Multi-Model Ecosystem and Licensing Considerations

The broader BFS repository encompasses a multi-model collection tailored for different base systems:

  • Qwen Family: Qwen-Image-2.1 (Head V1, Head V1.1, Body V1.0) and Qwen Image Edit 2509 / 2511
  • Additional Base Models: Flux 2 Klein, Krea 2, and experimental LTX-2 video in-context weights

When planning deployment into production or service workflows, developers must account for foundation model terms: Qwen-Image-2.1 base weights are distributed under a non-commercial research license, requiring careful review before commercial use.

Sources