Qwen-Image 2.1 Turbo BFS Face Swap - Enhanced ComfyUI One-Click Installer
The enhanced edition of my BFS (Best Face Swap) workflow for Qwen-Image 2.1 Turbo — it goes further than the free version to push resemblance and realism, and the one-click installer sets everything up for you.
What makes this version better: an automatic face close-up of your head reference fed in as a third image (so skin texture, freckles, eyelashes and iris detail carry over), a realism LoRA at 0.4 for believable skin, references fed in at 1536 for sharper detail, a detail-matching swap prompt and a built-in side-by-side preview.
Runs on the official Qwen-Image 2.1 Turbo model — 8 steps, CFG 1, no speed LoRA needed.
What the installer does: installs ComfyUI Portable, every custom node (ComfyUI Manager, ComfyUI-GGUF, rgthree, KJNodes, RES4LYF, ComfyUI-RMBG, Inpaint-CropAndStitch) and every model in the right folders. At the prompt, pick Turbo int8 (NVIDIA driver 580+/CUDA 13) or Turbo GGUF for low VRAM and older drivers. The face-segment model is pre-downloaded so your first run works offline.
Requirements: NVIDIA RTX 30/40/50-series GPU, ~8 GB VRAM for int8 or 6–8 GB with the GGUF, at least 25 GB free storage. Use responsibly — only swap faces of people who have given consent.
What you get
A one-click Windows installer for ComfyUI Portable, both Qwen-Image 2.1 workflows — Text-to-Image and Image Edit (I2I) — and the custom nodes they need: ComfyUI Manager, ComfyUI-GGUF and rgthree.
The download menu — pick your VRAM tier:
Base model: FP8 (int8, ~7 GB, higher quality) into diffusion_models/, or GGUF Q5_K_M (~4–7 GB by quant, low VRAM) into unet/ Text encoder: qwen3vl_8b int8 (standard) or w4a8 (extra low VRAM) into text_encoders/ VAE: qwen_image_2.1_vae_bf16 into vae/ — downloads with any choice
Optional turbo LoRAs into loras/: none, Viggle 6-step, Alibaba-PAI 4-step, or both to compare. They cut editing to 4–6 steps at cfg 1. They're consistently great for image editing, but text-to-image on them is hit-or-miss — use turbo for edits, and the standard model at full steps for reliable text-to-image. The matching free turbo edit workflow runs them.
System requirements
OS: Windows 10 or 11, 64-bit.
GPU: NVIDIA RTX 30XX / 40XX / 50XX. FP8: 12 GB. On a 16 GB card a 1024×1024 image takes about 1–2 minutes at 25 steps. GGUF + w4a8: runs on 6–8 GB and still produces good quality.
Disk: at least 30 GB free.
Handled for you: Git, 7-Zip and the correct PyTorch build are installed automatically.
Getting started
1. Put the install .bat in its own dedicated folder and double-click. Use a simple local path like C:\Qwen\ — avoid Desktop, Documents, and cloud-synced folders.
2. At the menu, pick the base model and text encoder for your card: 8 GB → GGUF + w4a8; 12 GB+ → FP8 + int8 for best quality. Add a turbo LoRA if you'll mostly be editing.
3. Load either workflow (Text-to-Image or Edit) and confirm the models are selected.
4. Start around 25 steps at 1024×1024.
5. For fast edits, run a turbo LoRA with the free turbo edit workflow at 4–6 steps and cfg 1.
Troubleshooting
Install fails or behaves oddly: check the install path first. A cloud-synced folder (OneDrive, Dropbox) or a deep path with spaces is the most common cause. Move to C:\Qwen\ and re-run.
Out of memory: choose the GGUF base and the w4a8 text encoder at the download menu — that's the 6–8 GB tier.
Text-to-image comes out smeared or incomplete: check whether a turbo LoRA is active. They're reliable for editing but hit-or-miss for text-to-image — turn it off and run the standard model at full steps.
A model isn't showing in a loader: the FP8 base lives in diffusion_models/ and the GGUF base in unet/, and they're loaded by different loader nodes — select the one that matches what you downloaded.
Licence & credit
Qwen-Image 2.1 is by the Qwen team (Alibaba). ComfyUI support is by Comfy-Org, the GGUF quants by Abiray, and the turbo LoRAs by Viggle and Alibaba-PAI. This product is the installer and workflow packaging — it automates the setup of their work into a one-click process.
Licence note: Qwen-Image 2.1 is released under the Qwen Research License — non-commercial use only. Unlike the earlier Apache-licensed Qwen-Image release, commercial use requires a separate licence from Tongyi Lab (Alibaba). The GGUF quants and turbo LoRAs are derived from it and carry the same licence. Read the terms on the model's Hugging Face page before any commercial use.