What is Qwen-Image 2.1?
Qwen-Image 2.1 is the Qwen team's most powerful open-source image model, released on September 20, 2026. It unifies text-to-image generation and image editing in a single 7B-parameter model (32 Single-Stream DiT layers in the visual generation component), balancing quality, inference efficiency and versatility. On October 9, 2026 the team also released Qwen-Image-2.1-Turbo, a distilled checkpoint that generates and edits in just 8 denoising steps.
Key Capabilities
- Unified generation and editing — create images from text, edit existing photos, and extract subjects, all in one model.
- Native transparency (RGBA) — generate regular or transparent images from a prompt, edit transparent layers directly, and cut subjects out of photographs without a separate tool.
- Multi-reference editing — up to 10 reference images for multi-subject composition, with local edits specified via circles, painted annotations or masks; preserves identity for people and products.
- Better textures and typography — improved text rendering, portrait lighting and fine details.
- Efficient — mixed-granularity attention and prefix KV cache reuse keep the 7B model fast on modest hardware.
How to Run It
- Diffusers — day-0 support via
QwenImage21Pipeline(bf16, CUDA). - ComfyUI — native day-0 support with compatible weights at Comfy-Org/Qwen-Image-2.1 and ready-made text-to-image and editing workflows.
- ModelScope / DiffSynth-Studio — generation plus LoRA training.
- Acceleration — SGLang (prefix caching, Cache-DiT, CUDA graphs, parallelism) and LightX2V provide day-0 inference acceleration.
- Hosted APIs — Qwen-Image-2.1 Pro and Turbo are officially available through Alibaba Cloud Model Studio.
Weights are on Hugging Face and ModelScope; source and docs live on GitHub.
License Note
Qwen-Image 2.1 ships under the Qwen Research License Agreement — permissive for research and evaluation, with restrictions on commercial use. Check the license terms before shipping it in a product; for hosted commercial generation, the official APIs are the compliant route.
How It Compares
- vs FLUX.1 — FLUX remains a strong open-weights aesthetic baseline, but editing workflows and transparency typically require separate tooling; Qwen-Image 2.1 unifies generation, editing and RGBA in one checkpoint.
- vs Midjourney / Ideogram — those are closed hosted products (Ideogram notably strong at text rendering); Qwen-Image 2.1 is the option you can run and fine-tune locally.
- Turbo vs base — same 7B architecture, but Turbo needs only 8 denoising steps (CFG=1, saved sampling schedule), trading a little flexibility for 5x-fewer-step speed.
Frequently Asked Questions
Is Qwen-Image 2.1 free?
The weights are free to download and run for research purposes under the Qwen Research License. Commercial use is restricted — review the license or use the official APIs.
What is Qwen-Image-2.1-Turbo?
A distilled version of the same 7B model released on October 9, 2026, doing generation and editing in 8 denoising steps with the recommended sampling schedule baked into the checkpoint.
Can it edit transparent images?
Yes — native RGBA support means you can generate transparent images from prompts and edit transparent layers (e.g. change a subject's expression while keeping the transparent background).
Findry coverage
- Qwen-Image 2.1 guide — Turbo 8-step update, local setup, license notes, and the FLUX comparison
Qwen-Image 2.1 Turbo: 8-Step Ultra-Fast Diffusion
In October 2026, the Qwen team released Qwen-Image 2.1 Turbo, an accelerated distilled variant engineered for real-time generative workflows and edge deployment.
Key Turbo Enhancements:
- 5x Faster Generation — High-fidelity 1024x1024 photorealistic output in just 8 sampling steps, down from 40-50 steps on the base model.
- ComfyUI & GGUF Support — Native ComfyUI workflow nodes and quantized GGUF weights, running smoothly on consumer 12GB/16GB VRAM GPUs.
- Zero Loss in Text Rendering — Retains full bilingual Chinese/English typography and multi-line poster composition capabilities.
- Instant Photo Editing — Sub-second localized inpainting, subject swaps, and reference-guided style transfer.
