Qwen Image 2.1 Turbo

Qwen Image 2.1 Turbo is Qwen's accelerated image-generation and editing checkpoint. Start with the exact Turbo model and its matching workflow, then decide whether its license and hardware requirements fit your task.

Official model card · ComfyUI Turbo workflow · Read the license

Independent directory guide. Sources checked October 23, 2026. Varwo has not run this model, measured speed or generated the examples below.

Check the research license before choosing a workflow

The checkpoint's Qwen Research License limits use of the materials to non-commercial research or evaluation. Commercial use requires a separate license. Downloadable weights do not establish permission to offer a commercial generator.

This listing links to upstream resources; Varwo does not host the weights or provide an inference service. A hosted provider has its own access, billing and service terms to check before use.

What Turbo changes

The model card identifies a 7B visual-generation architecture with an eight-step denoising schedule. The checkpoint loads through QwenImage21Pipeline; it supports creating an image from text and editing with a reference image.

The upstream project separately links local integrations and Alibaba Cloud Model Studio APIs. An API listing, a community demo and a local checkpoint are different access routes. Do not assume they share a free allowance, identical parameters or the same deployment terms.

Start with the ComfyUI Turbo template

Comfy Org publishes separate Turbo text-to-image and Turbo image-editing templates. Choose the one for your task rather than importing an unrelated Qwen Image workflow.

1. Update ComfyUI using its supported installation route, then open the chosen template.
2. Follow the model links and filenames in that template and the Comfy Org card. It documents diffusion weights under models/diffusion_models/, the encoder under models/text_encoders/ and the VAE under models/vae/.
3. Use either the documented full Turbo checkpoint or the documented base-plus-Turbo-LoRA route. They are alternatives; extra weights are not automatically required for both.
4. Begin with the published Turbo settings: eight steps and CFG 1. Record the exact template and selected files with your result.

These are source-based preparation steps, not a configuration installed or tested by Varwo.

Using Diffusers instead

Follow the model card's current installation section. It calls for source Diffusers support for the saved sampling schedule; a generic package-install snippet may lag that requirement. Let the checkpoint supply its recommended schedule initially. The card notes that changing num_inference_steps alone does not replace it.

Keep a short run record: model revision, Diffusers revision, device, precision, resolution, seed and whether setup/model loading is included in timing. Eight denoising steps are not a measured latency or a GPU-memory guarantee.

A first prompt you can evaluate

Use this original prompt as a small typography test:

Create a square poster for a fictional community seed exchange.
Cream background, one dark-green botanical line drawing, generous margins.
Show exactly two lines of text: "SEED SWAP" and "SATURDAY 10 AM".
No other lettering, logos or decorative text.

Check exact spelling, missing or extra letters, alignment and whether the requested number of text lines survived. Change one instruction at a time. Save failed outputs too; a selected showcase image is not a repeatable success rate.

For an editing trial, start with your own non-sensitive image and ask for one change while naming what should stay fixed. Compare those unchanged areas before judging the whole picture.

Frequently asked questions

Is there a free generator on this page?

No. This is a directory and setup guide with upstream links. Local execution needs your own compute; hosted services set their own limits and charges.

Can I use the research weights in a commercial product?

The published license requires separate commercial permission. Review the actual license and intended service before deploying it.

Is this the same as MOSS-Image?

No. They are different model families with different files and workflows. See the MOSS-Image overview and MOSS ComfyUI guide for that model.