Zwaetschge's Qwen Image 2.1 Turbo Workflows

# Qwen-Image 2.1 Turbo — 8-Step Text-to-Image

An English ComfyUI workflow for the full Qwen-Image 2.1 Turbo checkpoint, with a clear layout, built-in setup instructions, and optional LoRA support.

### Features

- INT8 ConvRot model for lower memory use.

- Official eight-step sigma schedule, Euler sampler, and CFG 1.

- Aspect-ratio and megapixel controls; default 2 MP portrait.

- Optional LoRA loader, bypassed by default.

- Built-in ComfyUI nodes only—no custom-node pack required by the graph.

- Model download links inside the workflow.

- PNG output with embedded workflow metadata.

### Setup

Update ComfyUI and its frontend to versions supporting Qwen-Image 2.1 and native INT8 ConvRot.

Download these files from [Comfy-Org/Qwen-Image-2.1](https://huggingface.co/Comfy-Org/Qwen-Image-2.1):

| File | Folder |

|---|---|

| qwen_image_2.1_turbo_int8_convrot.safetensors | models/diffusion_models/ |

| qwen3vl_8b_int8_convrot.safetensors | models/text_encoders/ |

| qwen_image_2.1_vae_bf16.safetensors | models/vae/ |

Load the workflow JSON, select the files in their loaders, write your prompt, choose the canvas, and click Run.

### Why the custom schedule?

Setting a regular Euler/simple KSampler to eight steps does not reproduce the Turbo schedule. This workflow uses the exact noise levels specified in the [official Turbo configuration](https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo/blob/main/model_index.json).

Keep the supplied schedule, Euler sampler, and CFG 1 unchanged. Use the full Turbo checkpoint without an additional acceleration or Viggle Turbo LoRA.

### Optional LoRAs

Choose an installed Qwen-Image 2.1 LoRA, then select the LoRA node and press Ctrl+B to enable it. Compatibility and useful strengths vary with Turbo; no LoRA is needed for the default workflow.

### Testing

The workflow was inspected and successfully run in ComfyUI 0.38.2, frontend 1.53.6, at 1184×1760.

I also tested generation at 12 megapixels, achieving surprisingly good results in approximately 200 seconds. This is a result from my own setup; generation time and memory requirements will vary with hardware and settings.

The included example used the locally installed nynxz INT8 ConvRot Turbo quant. The public JSON selects the official Comfy-Org filename for portability. That official quant was not GPU-tested during this release pass, so identical pixels are not guaranteed.

The included example used the locally installed nynxz INT8 ConvRot Turbo quant. The public JSON selects the official Comfy-Org filename for portability. That official quant was not GPU-tested during this release pass, so identical pixels are not guaranteed.

Credits to the Qwen team and ComfyUI / Comfy-Org. This is a community workflow, not an official Qwen release.