Turbo + Upscale for Qwen Image 2.1

This free ComfyUI workflow draws a picture from your prompt in 4 steps with Viggle's turbo LoRA, at 832 x 1248. A detail pass then makes it 1664 x 2496 and draws 4 more steps over it at denoise 0.35. Hair, knit and skin get real texture, and the face stays the same.

The "upscale on / off" switch skips the detail pass when you only want a quick look.

What you need

No custom nodes. Tested on ComfyUI 0.37.4. If a node shows red, update ComfyUI: older builds do not have the Qwen Image 2.1 nodes.

• qwen_image_2.1_int8_convrot.safetensors (7.3 GB) in models/diffusion_models
• qwen3vl_8b_int8_convrot.safetensors (9.4 GB) in models/text_encoders
• qwen_image_2.1_vae_bf16.safetensors (676 MB) in models/vae
• Qwen-Image-2.1-viggle-turbo-4step-lora-r64.safetensors (340 MB) in models/loras

When you open the workflow, ComfyUI offers to download all four into the right folders.

First picture

1. Drag the workflow onto the ComfyUI canvas.
2. Write what you want in the green "Your prompt" box. Full sentences work best: who, what they wear, where, the light, the lens.
3. Press Run. The picture is saved in output/StableYogi.

Settings

4 steps, CFG 1, euler, beta. We rendered the same prompts and seeds on each setting and picked from the pictures. On the simple scheduler a faint second copy of the person showed up on 4 of 11 pictures; on beta, 0 of 11, at the same speed.

Memory

We ran it with the graphics memory held to 7 GB and the detail pass on, and it finished. The model files are bigger than an 8 GB card, so plan for 32 GB of system RAM. Out of memory? Set "upscale on / off" to false.

Licence

Built with Qwen. Qwen Image 2.1 and the Viggle turbo LoRA are for non-commercial use only. Qwen is licensed under the Qwen RESEARCH LICENSE AGREEMENT, Copyright (c) 2026 Hangzhou Tongyi Laboratory Technology Co., Ltd.

More free workflows: Community Resources on my portal.