Qwen-Image-2.1 on RTX PRO 6000: 15-Image Quality and Performance Test

On September 24, 2026, a test was conducted using the Qwen-Image-2.1 model on an NVIDIA RTX PRO 6000 Blackwell Server Edition GPU. The test involved running 15 fixed-seed prompts through the official Diffusers pipeline, covering various subjects such as typography, product photography, architecture, landscapes, illustration, and a transparent asset. The model successfully completed all 15 requested samples without any out-of-memory or CUDA errors.

The average generation time for a 1-megapixel sample was approximately 9.375 seconds, with a median of 9.378 seconds. The model allocated an average peak GPU memory of 56.516 GiB for the 15-image run. The transparent sample demonstrated a real alpha channel, indicating the model's ability to handle complex textures.

The typography test revealed that the model correctly rendered all four requested words in different scenarios, including English neon signs, educational infographics, bilingual storefronts, and product labels. However, the model's ability to generate accurate instructional content was not fully tested, as only four observed samples were used.

The photographic output showcased the model's capability to generate natural-looking skin textures, warm lighting, believable fabric, and visible details such as dust and hands. However, the model occasionally strayed from the requested camera angle, as seen in the product photography example where the noodles were captured at a 45-degree angle instead of a top-down view.

Overall, the test demonstrated that Qwen-Image-2.1 on the RTX PRO 6000 Blackwell Server Edition can deliver high-quality outputs within a reasonable timeframe, making it a promising option for researchers and practitioners working with large GPU workstations.