BFS Head V1.1: Qwen Image 2.1 In-Context Face & Head Swap (Surgical Skin-Detail Ablation)

Full Resolution

Generation Specs RTX 4070

Sampler euler

Scheduler simple

Steps 12

CFG Scale 2.5

Seed 85244748616565

Target VRAM 12GB

VRAM 10GB - 12GB

GPU Hardware Compatibility

Optimal Ground Truth

Native GPU execution at full throughput with zero memory swapping.

Qwen Image 2.1 (DiT) Min 10GB VRAM 832×1248

BFS Head V1.1: Qwen Image 2.1 In-Context Face & Head Swap (Surgical Skin-Detail Ablation)

Open-source head-swapping architecture powered by Qwen Image 2.1 DiT and Alissonerdx's BFS V1.1 LoRA. Features surgical img_mlp.gate_up weight ablation to eliminate plastic skin softening while preserving 100% 3D head pose and complex ambient lighting integration.

Blueprint Summary RTX 4070 Verified

Reproducible ComfyUI workflow for BFS Head V1.1: Qwen Image 2.1 In-Context Face & Head Swap (Surgical Skin-Detail Ablation) using Qwen Image 2.1 (DiT) at 832×1248 resolution. Requires minimum 10GB VRAM with sampler euler and scheduler simple (12 steps). Includes 1-click terminal model sync and canvas JSON graph.

Node Graph Pipeline 10 Total Nodes

Verify in Resolver

01 Load Models

UnetLoaderGGUF

4 loaders (DiT, CLIP, VAE)

02 Adapters

LoraLoaderModelOnly

1 adapter

03 Conditioning

TextEncodeQwenImage21

1 prompt encodings

04 KSampler

KSampler

DiT latent denoising

05 Decode & Save

VAEDecode

Latent to pixel space

Model & Asset Setup 1-Click Script

Run in your ComfyUI root:

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

Positive 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>, 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

LoRA Adapter Stack 1 Adapters

RTX 4070 Calibrated Weights

01 bfs_head_v1.1_qwen_2.1.safetensors +1

Required Models 5 Models

Disk Space Required: 13.9 GB (5 models · DiT/Base: 8.5 GB · Text Encoder: 4.5 GB · LoRAs: 568 MB · VAE: 350 MB)

models/loras/ 248 MB HuggingFace

bfs_head_v1.1_qwen_2.1.safetensors Weight: 1.0 (Blocks 16-31 img_mlp.gate_up pruned to preserve natural skin micro-texture)

Get

models/loras/ 320 MB HuggingFace

p_qwen_image_2.1_8step_v0.1.safetensors Weight: 1.0 (Pruna AI 8-step Lightning acceleration LoRA)

Get

models/diffusion_models/ 8.5 GB HuggingFace

qwen_image_2.1_int8_convrot.safetensors Qwen Image 2.1 DiT INT8 quantized core

Get

models/text_encoders/ 4.5 GB HuggingFace

qwen3vl_8b_int8_convrot.safetensors Qwen3-VL Vision-Language Text Encoder (CLIP loader type: qwen_image)

Get

models/vae/ 350 MB HuggingFace

qwen_image_2.1_vae_bf16.safetensors Qwen Image 2.1 Official BF16 VAE

Get

Field Notes RTX 4070 Benchmark

Benchmark: 87.5s on RTX 4070 (10.6 GB VRAM peak, Euler simple, 12 steps, 832x1248)

Reverse-Engineered & Synthesized In-Context Swapping Architecture: Built upon Alibaba's Qwen Image 2.1 Diffusion Transformer (DiT) paired with Alissonerdx's breakthrough BFS Head V1.1 LoRA (248 MB). Unlike legacy embedding-pasting tools (InsightFace, ReActor, RoOP) that suffer from boundary seams and flat plastic skin, BFS performs native in-context latent regeneration. Key breakthrough: The V1.1 release executes surgical weight pruning on the second-half network (blocks 16-31), disabling 16 img_mlp.gate_up projections that previously caused artificial skin smoothing, thereby recovering 33% bilateral micro-texture (pores, freckles, stubble) without retraining. The early MLP blocks (0-15) and all 128 attention projections are byte-identical to V1, preserving full 3D head rotation, eye gaze, and subtle micro-expressions. Hardware ground truth verified on dedicated RTX 4070: 10.60 GB peak VRAM, 87.5s generation time for full resolution native in-context editing. Input Contract: Image 1 must be the Target Scene/Body (composition & lighting anchor), while Image 2 is the Reference Identity.

Frequently Asked Questions FAQ

What GPU and VRAM are required to run BFS Head V1.1: Qwen Image 2.1 In-Context Face & Head Swap (Surgical Skin-Detail Ablation)?

This workflow requires a minimum of 10GB VRAM (recommended 12GB VRAM). Tested and verified on NVIDIA GeForce RTX 4070 (12GB VRAM) at 832x1248 resolution.

How do I resolve missing custom nodes for this workflow?

You can drop the workflow JSON into our client-side Missing Node Auto-Resolver at https://comfy.yellorn.com/resolve/ to detect missing nodes and generate install commands, or run the 1-click terminal setup script provided below.

What conditioning and VAE architecture does BFS Head Swap require?

BFS Head Swap V1.1 uses Qwen Image 2.1 DiT with dual VAE conditioning and in-context facial geometry alignment. It performs surgical skin-detail ablation without requiring external face-swapping models.