Model guide

Qwen-Image-2.1-Turbo

Qwen-Image-2.1-Turbo is an accelerated checkpoint of Qwen-Image-2.1, released on October 9, 2026. It keeps the same 7B visual generation architecture and finishes an image in 8 denoising steps, where the base checkpoint defaults to 40. One model covers both text-to-image generation and image editing.

Turbo vs the Qwen-Image-2.1 checkpoint

Turbo is not a smaller model. It is the same network with a sampling schedule tuned for fewer steps.

Item	Qwen-Image-2.1-Turbo	Qwen-Image-2.1
Denoising steps	8, from the schedule saved in the checkpoint	40 by default
Changing the step count	Setting num_inference_steps alone does not override the saved schedule. Pass sigmas to experiment.	Set num_inference_steps
Architecture	7B visual generation component, 32 single-stream DiT layers	Same
Default resolution	2048 level (2048 × 2048 at 1:1)	Same
Tasks	Text-to-image and image editing	Same
License	Qwen Research License Agreement	Same

What Turbo keeps from Qwen-Image-2.1

Text to image

Write a prompt and get a 2K image in one of seven preset aspect ratios.

Reference-based editing

Edit from up to 10 reference images, for example turning a sketch into a photorealistic render.

Transparent output

The 64-channel RGBA autoencoder can generate images with a real alpha channel for stickers and cutouts.

Text and layout

Posters, study notes, and UI-style layouts with long passages of readable text are part of the official Turbo showcase.

Run Qwen-Image-2.1-Turbo with Diffusers

The official model card loads Turbo with QwenImage21Pipeline. You need a CUDA GPU and a Diffusers build from source.

Install the dependencies
Install a CUDA build of PyTorch, then Diffusers from GitHub and transformers 5.17.0 or newer. The release build of Diffusers may not yet support schedules configured inside a pipeline, which Turbo needs.
Load the checkpoint
Load Qwen/Qwen-Image-2.1-Turbo in bfloat16. The 8-step schedule is read from the checkpoint, so you do not configure a scheduler.
Generate at a preset size
Pick a width and height from the table below and keep use_kv_cache on. Text and reference images are encoded once and reused across all 8 steps.
Edit an existing image
Pass the source image as image= to the same pipeline call. The official example converts it to RGBA first.

pip install git+https://github.com/huggingface/diffusers.git
pip install "transformers>=5.17.0" accelerate pillow

import torch
from diffusers import QwenImage21Pipeline

pipe = QwenImage21Pipeline.from_pretrained(
    "Qwen/Qwen-Image-2.1-Turbo", dtype=torch.bfloat16
).to("cuda")

image = pipe(
    prompt="Editorial poster for a night market, bold headline NIGHT MARKET, gold ink on charcoal paper",
    width=2528,
    height=1696,
    use_kv_cache=True,
    generator=torch.Generator("cpu").manual_seed(42),
).images[0]
image.save("turbo.png")

Resolution presets

Turbo uses the same presets as Qwen-Image-2.1.

Aspect ratio	Width × height (px)
1:1	2048 × 2048
4:3	2400 × 1792
3:4	1792 × 2400
3:2	2528 × 1696
2:3	1696 × 2528
16:9	2752 × 1536
9:16	1536 × 2752

Local Turbo or the online workspace

Run Turbo locally when

You have a CUDA GPU, need many drafts in a row, or want every file to stay on your own machine.

Use the online workspace when

You have no GPU, want to paint a mask instead of writing code, or want 1K and 2K presets with a saved history.

Before you rely on Turbo

The weights are published under the Qwen Research License Agreement. Read it before any commercial use.
Only the saved 8-step schedule has been evaluated. Custom sigmas are for experiments.
The model card does not list minimum VRAM. Test on your own GPU before you plan batch jobs.