BFS Head V1.1: Qwen Image 2.1 In-Context Face & Head Swap Guide (Setup & Download)

Overview: Why BFS Head V1.1 Over Legacy Face Swapping

Traditional face-swapping pipelines (InsightFace, ReActor, RoOP) operate by detecting a 2D bounding box, extracting facial identity embeddings, and warping a pre-rendered face onto the target body. This legacy approach consistently suffers from three critical flaws:
1. Unnatural Neck Seams: Pasting pixels across different skin tones creates noticeable boundary artifacts.
2. Lighting Mismatch: Cropped face embeds do not inherit ambient lighting, room bounces, or directional shadows from the target scene.
3. Plastic Skin Smoothing: Over-aggressive anti-aliasing erases skin pores and natural micro-textures.

BFS Head V1.1 (developed by Alissonerdx / NRDX) solves this by operating directly inside Alibaba Qwen Image 2.1 Diffusion Transformer (DiT). Instead of post-processing masks, Qwen Image 2.1 processes both images in context within its cross-attention layers. The identity from <image2> is natively hallucinated onto the pose and background of <image1> with physically consistent shadows, head tilt, and micro-expressions.

---

Key Breakthrough in V1.1: Surgical Weight Ablation

The initial BFS V1 suffered from excessive skin smoothing. The V1.1 release fixes this through mathematical ablation:
- 16 Pruned MLP Projections: In the second half of the network (blocks 16 to 31), 16 img_mlp.gate_up projection weights were ablated.
- 33% Micro-Texture Recovery: Disabling these late MLP layers preserves natural skin pores, fine wrinkles, and stubble without requiring full model retraining.
- Identical 3D Pose Geometry: Blocks 0 to 15 and all 128 attention projections remain byte-identical to V1, perfectly preserving head angle, eye gaze, and micro-expressions.

---

Hardware Benchmarks & Generation Telemetry

Tested and verified on a dedicated workstation with an NVIDIA GeForce RTX 4070 (12GB VRAM):
- Diffusion Core: qwen-image-2.1-UC-Q4_K_M.gguf (4.39 GB)
- Text Encoder: qwen3vl_8b_w4a8_heretic.safetensors (6.01 GB, CLIPLoader type: qwen_image)
- VAE: qwen_image_2.1_vae_bf16.safetensors (644 MB)
- LoRA Adapter: bfs_head_v1.1_qwen_2.1.safetensors (Weight: 1.0)
- Sampler: Euler | Scheduler: Simple
- Steps: 12 | CFG Scale: 2.5 | Seed: 85244748616565
- Resolution: 832x1248 (Preserves native ~2:3 portrait aspect ratio)
- Peak VRAM: 10.60 GB (Fits safely on 12GB cards)
- Sampling Latency: 87.5 seconds

---

Step-by-Step Setup Guide

Method 1: 1-Click Automated Synchronization

To download all required model weights, text encoders, VAE, and BFS LoRA into your ComfyUI root directory in a single automated pass:

Linux / macOS:
```bash
curl -fsSL https://comfy.yellorn.com/api/scripts/bfs-head-swap-v1-1-qwen-image-2-1.sh | bash
```

Windows PowerShell:
```powershell
irm https://comfy.yellorn.com/api/scripts/bfs-head-swap-v1-1-qwen-image-2-1.ps1 | iex
```

Method 2: Manual Directory Placement

1. Diffusion Model: Place qwen-image-2.1-UC-Q4_K_M.gguf inside ComfyUI/models/unet/
2. Text Encoder: Place qwen3vl_8b_w4a8_heretic.safetensors inside ComfyUI/models/clip/
3. VAE: Place qwen_image_2.1_vae_bf16.safetensors inside ComfyUI/models/vae/
4. LoRA: Place bfs_head_v1.1_qwen_2.1.safetensors inside ComfyUI/models/loras/

---

The In-Context Prompt Contract

In the Qwen Image 2.1 architecture, <image1> serves as the composition and lighting anchor, while <image2> provides identity reference:

### Positive Prompt:
```text
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>, but maintain the aspect ratio of the head from <image1>, maintain the direction of the eye, head rotation, skin color, micro expressions from <image1>. high quality, sharp details, 4k
```

### Negative Prompt:
(Leave empty)

---

Download Pre-Wired ComfyUI Canvas Workflow

You do not need to route cables manually or fix missing node errors. A clean 10-node workflow JSON pre-wired for this architecture is available on Comfy Lab:

👉 Download Pre-Wired Workflow JSON on Comfy Lab

Simply drag and drop the downloaded .json file directly into your ComfyUI window, select your target scene as Image 1 and your identity face as Image 2, and click Queue Prompt.