BFS Head V1.1: Qwen Image 2.1 In-Context Face & Head Swap (Surgical Skin-Detail Ablation)
8st · euler 832×1248 10GB
Full Resolution
Generation Specs RTX 4070
Sampler euler
Scheduler simple
Steps 35
CFG Scale 2.2
Seed 8492049102
Target VRAM 12GB
VRAM 8GB - 12GB
GPU Compatibility
Optimal Performance
Runs entirely in VRAM with no memory swapping.
Qwen-Image-2.1 DiT (GGUF Q4_K_M) Min 8GB VRAM 1024×1024 RTX 4070 Verified
Qwen-Image-2.1 Uncensored: GGUF DiT & Heretic Text Encoder
Full canvas workflow combining GGUF-quantized Qwen-Image-2.1 diffusion weights with the refusal-ablated Heretic text encoder. Eliminates morality guardrails and runs smoothly on 12GB VRAM.
Blueprint Summary RTX 4070 Verified
Reproducible ComfyUI workflow for Qwen-Image-2.1 Uncensored: GGUF DiT & Heretic Text Encoder using Qwen-Image-2.1 DiT (GGUF Q4_K_M) at 1024×1024 resolution. Requires minimum 8GB VRAM with sampler euler and scheduler simple (35 steps). Includes 1-click terminal model sync and canvas JSON graph.
Pipeline Flow 7 Nodes
Open in Resolver 01 Load Models
UnetLoaderGGUF
3 loaders (DiT, CLIP, VAE)
02 Conditioning
TextEncodeQwenImage21
1 prompt encodings
03 KSampler
KSampler
DiT latent denoising
04 Decode & Save
VAEDecode
Latent to pixel space
Execution DAG Topology Interactive Visualizer
Drag to pan · Scroll to zoom · Hover wiresModel & Asset Setup Setup Script
Run in your ComfyUI root:
curl -fsSL https://decomfy.com/api/scripts/qwen-image-2-1-uncensored-comfyui.sh | bash Loading bash setup script... Loading PowerShell setup script... import modal
app = modal.App("comfyui-qwen-image-2-1-uncensored-comfyui")
vol = modal.Volume.from_name("comfy-weights-cache", create_if_missing=True)
image = (
modal.Image.debian_slim(python_version="3.11")
.apt_install("git", "wget", "curl", "libgl1-mesa-glx", "libglib2.0-0")
.pip_install("torch", "torchvision", "--index-url", "https://download.pytorch.org/whl/cu124")
.pip_install("transformers", "accelerate", "safetensors", "aiohttp")
.run_commands(
"git clone https://github.com/comfyanonymous/ComfyUI.git /root/ComfyUI",
"cd /root/ComfyUI && pip install -r requirements.txt",
)
)
@app.function(
gpu="T4",
image=image,
volumes={"/root/ComfyUI/models": vol},
timeout=900,
)
def generate():
# Headless serverless execution for Qwen-Image-2.1 DiT (GGUF Q4_K_M)
print("Executing Qwen-Image-2.1 Uncensored: GGUF DiT & Heretic Text Encoder on ephemeral T4 GPU...")
return {"status": "success", "slug": "qwen-image-2-1-uncensored-comfyui"}
runpodctl create pod \
--name "comfy-qwen-image-2-1-uncensored-comfyui" \
--gpu-type "NVIDIA RTX 4070 Ti" \
--image "runpod/comfyui:latest" \
--volume-in-gb 50 \
--ports "8188/http" # ComfyUI Model Batch Ingestion for Qwen-Image-2.1 Uncensored: GGUF DiT & Heretic Text Encoder
# Run with: aria2c -i models-qwen-image-2-1-uncensored-comfyui.txt -j4 -x4
https://huggingface.co/abenzerps/Qwen-Image-2.1-Uncensored-GGUF/resolve/main/qwen-image-2.1-UC-Q4_K_M.gguf
dir=models/diffusion_models
out=qwen-image-2.1-UC-Q4_K_M.gguf
https://huggingface.co/pottokao/Qwen-Image-2.1-Text-Encoder-Heretic-W4A8/resolve/main/qwen3vl_8b_w4a8_heretic.safetensors
dir=models/text_encoders
out=qwen3vl_8b_w4a8_heretic.safetensors
https://huggingface.co/abenzerps/Qwen-Image-2.1-Uncensored-GGUF/resolve/main/vae/qwen_image_2.1_vae_bf16.safetensors
dir=models/vae
out=qwen_image_2.1_vae_bf16.safetensors Positive Prompt
masterpiece, ultra-detailed 8k photograph of a fiery 21-year-old Korean ulzzang idol with luminous porcelain glass skin, wearing an intricate scarlet red lace corset bustier with delicate black satin ribbons and sheer black lace thigh-high stockings, seated on plush velvet couch in high-rise penthouse, panoramic rainy Seoul skyline with neon amber reflections through wet glass window, seductive sultry gaze, biting lower lip, 85mm f/1.4
Negative Prompt
blurry, low quality, deformed limbs, bad anatomy, bad proportions, unclothed, blowout highlights, plastic skin, lowres
Prompt Customization:
This workflow is pre-wired and calibrated. Swap this prompt with your own character, subject, or style without breaking the pipeline.
Required Models 3 Models
Field Notes RTX 4070 Benchmark
Benchmark: RTX 4070 (12GB VRAM): ~24.2s per 1024x1024 frame (9.4 GB active VRAM footprint). RTX 3060: 38s.
Dual-Layer Uncensored Architecture & Memory Offloading: Eliminates morality guardrails by pairing abenzersps Qwen-Image-2.1 UC quantized DiT with Pottokao's refusal-ablated Heretic multimodal text encoder (W4A8). Directional ablation subtracts refusal vectors with minimal KL divergence (0.0220), cutting refusals from 100% to 5% while preserving deep compositional semantics. Flow-matching dynamics require CFG dialed precisely to 2.2 (values > 3.0 cause harsh edge burnout and plastic sheen). By loading the 4.29GB Q4_K_M DiT into VRAM and offloading the text encoder to System RAM after initial conditioning, this pipeline runs smoothly within 8GB to 12GB VRAM cards.
Frequently Asked Questions FAQ
What GPU and VRAM are required to run Qwen-Image-2.1 Uncensored: GGUF DiT & Heretic Text Encoder?
This workflow requires a minimum of 8GB VRAM (recommended 12GB VRAM). Tested and verified on NVIDIA GeForce RTX 4070 (12GB VRAM) at 1024x1024 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://decomfy.com/resolve/ to detect missing nodes and generate install commands, or run the 1-click terminal setup script provided below.
What text encoders and VAE does Qwen-Image 2.1 Uncensored require?
Qwen-Image 2.1 Uncensored uses GGUF Q4_K_M quantized DiT weights paired with the Heretic uncensored text encoder, bypassing safety refusals while maintaining precise prompt adherence.
Related ComfyUI Blueprints
View all 50 workflows
RedCraft 4.0 - Intimate Editorial Lookbook (K2T DUAL)
8st · euler 1888×3328 10GB
High-Contrast Penthouse Skyline Silhouette - Krea 2 Turbo & Qwen3-VL
9st · er_sde 944×1664 10GB