MODEL PROFILE · Qwen / Alibaba
Qwen-Image-2.1
A downloadable image model with documented RGBA generation and editing. Evaluate transparent edges and reference fidelity separately.
Source reviewed: · Image model
Provider documentation summary. No hands-on results or independent ranking are claimed.
Documented facts
The following fields come from the official provider source, reviewed on the date above.
- Identity and access
- The official Qwen/Qwen-Image-2.1 repository provides weights and Diffusers instructions. This catalog addition is not a new-release announcement.
- Inputs and outputs
- The card describes text-to-image generation, image editing and transparent RGBA output.
- Reference and edit controls
- Qwen documents up to 10 reference images, with circles, painted annotations or separate masks for local changes.
- Documented dimensions
- Examples use 2048 × 2048 pixels; the aspect-ratio table includes landscape 2752 × 1536 and portrait 1536 × 2752. These are documented configurations, not measured quality guarantees.
- Architecture qualifier
- The stated 7B parameter count applies to the visual generation component. It is not a verified memory requirement for a complete deployment.
- License
- The repository identifies the Qwen Research License Agreement. Read its terms before choosing a deployment; downloadable weights do not establish unrestricted commercial permission.
- Pricing and API
- Hosted API rates and availability are not verified here. Local inference still has hardware and operating costs.
Qwen-Image 2.1 versus the 3.1 roadmap
Alibaba’s September 22, 2026 announcement describes Qwen-Image 3.1 as planned for later in 2026, with transparent backgrounds and image editing among its advertised capabilities. That is a roadmap statement, not proof of public access. The downloadable repository documented above is Qwen-Image-2.1. Do not transfer its reference-image limits, dimensions, architecture or license to 3.1. We have not verified a 3.1 checkpoint, public API, price or exact launch date. Alibaba’s dated Qwen-Image 3.1 roadmap.
What this means for an evaluation today
If you need to evaluate downloadable weights now, record the exact 2.1 repository revision. If you are waiting for 3.1, treat its proposed capabilities as a future requirement to verify, not a measured improvement over 2.1. Transparency is already documented for 2.1, so the feature name alone does not establish a quality upgrade. Check alpha edges, unchanged regions and reference fidelity on the actual version you use. Qwen-Image-2.1 model card.
The practical decision
For a sticker, product cutout or layered composition, a plausible subject is only part of an acceptable result. Our suggested review is to place the saved RGBA output over both a light and a dark background, then inspect fine hair, semitransparent material and boundary halos. Keep that assessment separate from whether a requested edit preserved the subject. For a multi-reference task, write down which reference supplies each attribute before reviewing the output; otherwise an attractive result can conceal a missed instruction. These are evaluation suggestions, not findings from a DualView test.
What to record in your own comparison
- Keep the exact weight revision, inference library version, prompt, seed, step count, dimensions and input files with every output.
- Preserve the alpha channel in the saved file. A flattened preview cannot establish that the delivered asset has useful transparency.
- Review requested changes and untouched regions separately. Record rejected attempts alongside accepted outputs when estimating cost.
Limitations and unknowns
DualView has not generated samples or measured this model’s speed, memory use, typography, identity retention or quality. The model card establishes documented capabilities, not reliability on your assets. Hardware feasibility depends on the full configuration; a parameter count alone is insufficient. No provenance-matched example or independent ranking is presented.
Compare your saved results
Use the DualView editor to inspect files you already have. The accepted-output calculator uses your own rates and counts. Neither link runs a model or purchases generation.