How to Run Qwen Image 2.1 Turbo in ComfyUI
Qwen-Image-2.1 runs locally in ComfyUI, and thanks to community GGUF quantizations it fits on 8–12 GB graphics cards. This guide walks through the parts that trip people up: which files you need, which folder each one goes in, which GGUF loader understands the new qwen_image21 architecture (the original ComfyUI-GGUF node throws "Unknown model architecture"), and how to set up the 8-step Turbo checkpoint. VRAM figures here come from community model cards and tests and are labeled that way, because we haven't benchmarked every card. Keep in mind the weights are released under the Qwen Research License, which restricts commercial use. If you just want images and don't want to spend an evening on setup, the same model runs in our online generator, where the first image is free.
What you need
| VRAM | Suggested file | Source |
|---|---|---|
| 8 GB | Q4_K_M GGUF | Community report (gatuwo) |
| 12 GB+ | Q8 GGUF | Community report (AlperKTS) |
Files to download and where they go
- GGUF or bf16 diffusion model →
models/diffusion_models/(orunet/) - Qwen3-VL 8B text encoder →
models/text_encoders/ qwen_image_2.1_vae_bf16.safetensors→models/vae/
Install a GGUF loader that supports 2.1
city96's original node shows "Unknown model architecture". Community forks (leejet, molbal) add support; check their READMEs for the current state.
Build the workflow
CLIPLoader with type qwen_image, Unet Loader (GGUF), KSampler with CFG 1.
Turbo-specific settings
8 steps, CFG 1. The best sampler and scheduler in ComfyUI is still being tested; we will update this section after our own runs.
Editing workflows
For editing, feed reference images into the conditioning.
Common errors and fixes
- "Unknown model architecture": switch to a loader fork that supports 2.1.
- Out of memory: drop to a smaller quant.
License note
The open weights are released under the Qwen Research License, which allows research and evaluation use only; commercial use needs a separate licence from Qwen.
No GPU? Try it online free: 2 images without sign-up. See pricing.
Don't want to install? Run it online
Building your own app? See the API guide. Or go straight to the generator.
FAQ
How much VRAM does Qwen Image 2.1 need?
Community reports: a Q4_K_M GGUF has been run on 8 GB, Q8 needs 12 GB or more. We have not benchmarked every card.
Which GGUF quant should I use?
Q4_K_M for 8 GB cards, Q8 if you have 12 GB or more.
Why do I get "Unknown model architecture"?
The original ComfyUI-GGUF node does not know the new qwen_image21 architecture. Use a fork that supports it.
Does Turbo work in ComfyUI?
Yes with the right loader; use 8 steps and CFG 1. We are still testing the best sampler and scheduler.
Is the "Uncensored GGUF" official?
No, it is a third-party conversion. We don't link to it.