Qwen Image 2.1

What it is

Qwen-Image-2.1-Turbo is an accelerated checkpoint of Qwen-Image-2.1: same 7B architecture, ~8-step denoising, with a recommended sampling schedule baked in. Default CFG=1; prefix KV caching reuses text and reference-image context across steps.

• Text-to-image and image editing share one pipeline.
• Aspect-ratio presets match the base (e.g. 1:1 → 2048×2048, 2:3 → 1696×2528).
• Model-card showcase covers portraits, motion, transparency, typography/posters, single-image transforms, multi-reference and interior composition.

Who made it

Released by the Qwen team on Hugging Face as Qwen/Qwen-Image-2.1-Turbo, finetuned from Qwen/Qwen-Image-2.1. Licensed under the Qwen Research License Agreement (HF license: other). Use the official feedback form on the card to reach the research team.

What it is for

From the model card:

• Text-to-image with detailed prompts (example portrait 1696×2528).
• Image editing / reference transforms (example 2048×2048 yacht sketch → photoreal photo).
• Multi-reference / interior layouts as shown in the showcase.

Needs Diffusers with pipeline-configured sampling sigmas (PR #14950); install from Diffusers git as documented.

How to try

Install CUDA PyTorch, then:

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="Your prompt here",
    width=1696,
    height=2528,
    use_kv_cache=True,
    generator=torch.Generator("cpu").manual_seed(42),
).images[0]
image.save("t2i.png")

For editing, pass image=... plus an edit prompt. The 8-step schedule loads with the checkpoint; num_inference_steps alone does not override it. See the card for full examples and aspect-ratio table.

vs base / quantized cousins

vs Qwen-Image-2.1 base: Turbo is the speed path—fewer steps, saved schedule, KV cache. Quality/prompt fidelity details belong on the official blog and cards.

Quantized / community forks: HF trending may list Uncensored-Quantized and similar third-party packs. They are not this page’s subject; search HF if you need smaller footprints, and verify license/safety differences yourself. No separate quantized landing page in this batch.