BFS Head Swap Adds Qwen-Image 2.1 Face Replacement

One of the eight head-swap examples published with the release.

The collection is one of the more heavily used face-editing repositories on Hugging Face, with roughly 889 likes and 131,000 downloads at the time of writing, and it comes from the same author behind the H3 sharpness LoRA and the H3 VFX Edit pack covered earlier in September. Until now its image-side weights targeted Qwen Image Edit 2509 and 2511 and FLUX.2 Klein. The new adapter targets the base Qwen-Image 2.1 model instead, so it does not replace the older Qwen Image Edit releases.

What the update adds

The model card was also reorganized into version guides, so each family now documents its own base model, image order and starting prompt instead of sharing one generic description.

Running the Qwen-Image 2.1 head swap

The workflow takes two images, and the order is part of the contract:

• Image 1 is the base image. Its body, pose, composition, lighting and environment provide the foundation for the result.
• Image 2 is the reference head. Its identity, hair, eye colour and nose structure are transferred.

The guide ships this starting prompt, which keeps the edit local to the head:

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

ComfyUI setup follows the usual LoRA path. Put bfs_head_v1_qwen_2.1.safetensors in ComfyUI/models/loras/, load the Qwen Image 2.1 diffusion model with its Qwen3-VL text encoder and matching VAE, then wire the target image to Image 1 and the reference head to Image 2. Start at LoRA strength 1.0 and only adjust after checking the identity, hair boundary, skin blending and expression transfer. Custom sizes in multiples of 32 are supported, and the guide notes that larger outputs tend to hold facial detail better, since low resolutions can shift colour and softness.

A second head-swap example from the repository.

Experimental LTX-2 weights

The LTX-2 checkpoints are in-context LoRAs kept deliberately separate from the image releases: no LTX-2 workflow is bundled, and the documentation warns that the Qwen or FLUX input order does not carry over. They only apply inside an LTX-2 workflow that implements the matching in-context LoRA contract.

A third published example.

Limitations

The published limitations are specific: small faces, extreme angles, occlusion, hands over the face, hair boundaries and strong lighting changes all reduce fidelity. The guide also asks users to inspect the result for unintended changes to the body, clothing, background and text, and to avoid results involving public figures or people who have not given consent.

The rest of the collection covers Qwen Image Edit 2509 (head v1 to v4 plus a separate face adapter), Qwen Image Edit 2511 (v5, with merged rank 16 and rank 32 variants), FLUX.2 Klein 4B and 9B, and Krea 2 head v1.1 plus an experimental body swap.