Qwen-Image-2.1: Transparent Images and On-Device AI Editing
Published September 21, 2026
Alibaba's Qwen-Image-2.1 bundles image generation, photo editing and transparent backgrounds in one model — and a documented test shows it running locally on an RTX 4090. Up to ten reference images can be merged into one scene. But the headline numbers come with fine print on model size, benchmarks and licensing.
Transparency as a built-in feature
Qwen-Image-2.1 supports RGBA output, adding an alpha channel that controls per-pixel transparency. Mascots without a background, cut-out products or illustrations for web layouts can be generated directly as transparent images. Existing transparent images can be edited — Qwen demonstrates changed facial expressions and replaced lettering while preserving transparency — and objects in ordinary photos can be lifted out as transparent layers. The SGLang documentation also describes partially transparent transitions for soft edges; results still need a visual check.
Ten reference images for one scene
Up to ten reference images can feed a single edit. Manufacturer examples include group portraits, complete outfits and furnished rooms; for a virtual try-on, the model combines a person with separate shots of clothing, shoes and accessories. Changes can be constrained spatially using colored circles, painted areas or a separate mask that marks the intended edit without covering the original. Qwen also says faces stay more stable across edits, products keep their shape and labels, and typography, lighting and fine details have improved.
Compact generator, larger footprint
The widely cited seven billion parameters apply only to the image-generation component. The full system adds a Qwen3-VL encoder with eight billion parameters, and the complete download weighs in at 33.1 GB. Computation gets more efficient by reusing processed information about input images and editing instructions, so unchanged context does not have to be recomputed in every step — an advantage with several references.
Local run on an RTX 4090 — with conditions
The included SGLang documentation confirms a test on a GeForce RTX 4090 with 24 GB of VRAM: at 1,024 × 1,024 pixels and 40 sampling steps, the median response time was 18.68 seconds for generation and 21.68 seconds for editing, with a measured peak of 22.7 GiB of graphics memory. Parts of the encoder were offloaded to system RAM layer by layer. The figures were recorded after warm-up, including PNG output but excluding startup time — and they describe this specific configuration only. The runtimes do not transfer automatically to the also supported native 2K output.
Benchmark and license deserve a closer look
On the Qwen-Image-Bench, all-ai.de reports 60.28 points and the top spot among the compared open-weight models — but several proprietary models still rank ahead, and the benchmark comes from the Qwen ecosystem itself, making it an indicator rather than independent proof. The same precision applies to openness: the weights are available and tools such as Diffusers, ComfyUI and SGLang are supported, yet the repository lists its own "Qwen Research License Agreement", and commercial use reportedly requires separate permission. The "open source" label should therefore not be read as a promise of unrestricted use.
Sources
- Qwen-Image-2.1: Compact, Efficient, and Unified Image Creation – Qwen Team
- Qwen/Qwen-Image-2.1 – Model Repository – Hugging Face
- QwenLM/Qwen-Image-2.1 – GitHub
