AI Action Lab tested Qwen-Image-2.1 by generating a fictional iced coffee. The first result included unwanted text despite our instruction. After a prompt revision, the second result had no extra text visible in the screenshot we inspected. Here is what we did and what these two runs can tell us.
What is Qwen-Image-2.1?
Qwen announced Qwen-Image-2.1 on 20 September 2026 as a model for image generation and editing, including transparent images. These are capabilities described by the developer. Our test covered text-to-image generation only, not editing or transparent output. Official announcement
Three steps in the official demo
1. Open the official interactive demo and enter a description of the image you want. We asked for a fictional iced coffee in an unbranded glass, with no text or logos.
2. Check the settings. We used Speed 1024 and kept prompt enhancement off in both runs. Keeping these settings unchanged allowed us to observe the result after revising the prompt.
3. Generate the image and inspect it. Look for added writing or logos and details that differ from your description. A successful generation does not mean the result followed every instruction.
Create a clean square product photograph of a fictional iced coffee in a clear unbranded glass on a warm wooden table, soft morning window light, cream background, room for an English headline at the top. No text, no logos, no people.Get curated AI news by email
Three topics per day, with a Laos perspective. Unsubscribe anytime.
What changed between the two runs?
The first image included unwanted English and Chinese text, some of it nonsensical. The instruction to avoid text did not prevent this result.
For the second run, we removed the request to leave space for an English headline and explicitly forbade letters and typography. We found no extra text visible in the screenshot of that result. The change is an observation from these two runs, not proof that this prompt will always work.
What to know before using a result
This fictional coffee image is an AI-generated illustration, not a photograph of a real item for sale. Two runs are too few to establish a success rate, compare models or measure reliability. Review the complete job, including waiting and corrections, rather than judging only the first preview.
The published Qwen Research License permits research and evaluation of the model materials; commercial use requires a separate license. Model licensing, hosted-service terms and rights in generated output are separate questions. Check the applicable terms before commercial use. Published license
Get curated AI news by email
Three topics per day, with a Laos perspective. Unsubscribe anytime.