RTX PRO 6000 runs Qwen-Image-2.1 at 9.4 s per image, 52 s for native 2K

What was tested

According to a dev.to evaluation, a single NVIDIA RTX PRO 6000 Blackwell Server Edition was used to run 15 fixed-seed prompts through the official Diffusers pipeline (QwenImage21Pipeline) on September 24. The card reported 95.0 GiB of memory to PyTorch, running on driver 580.105.08 with PyTorch 2.11.0+cu128 and CUDA runtime 12.8. Every generation ran in BF16 with 40 inference steps and batch size 1, with no quantization, model sharding or CPU offload.

Qwen-Image-2.1 combines text-to-image generation, image editing, multilingual typography, native 2K output and RGBA transparency in one model, and the prompt set was chosen to exercise all of it: English and bilingual signage, portrait and product photography, architecture, landscape, illustration, a transparent asset and one direct 2048 × 2048 render. The author states that no prompt was rerun to obtain a more favorable output.

The numbers

• Fourteen images at roughly 1 megapixel averaged 9.378 seconds each (range 9.271 to 9.524 seconds, median 9.375 seconds), with average peak allocated GPU memory of 36.801 GiB.
• The single native 2048 × 2048 render took 52.021 seconds and peaked at 56.516 GiB.
• All 15 requests completed without an out-of-memory or CUDA error.
• Warm-cache pipeline loading took 4.427 seconds, and total generation time across the set was 183.317 seconds.
• At the measured sequential rate, the author estimates roughly 384 images per GPU-hour at 1 MP, before orchestration, file I/O, safety review, retries or loading overhead.

The memory figures are PyTorch peak allocated memory per generation, not reserved memory or whole-system consumption.

Native 2K is practical, but costly

Direct 2K generation worked without a separate upscaler and stayed well inside the card's capacity, but it cost about 1.54× the memory and 5.55× the time of the 1 MP group for roughly four times the pixels. The dev.to evaluation found the 2K result coherent and richly textured at normal viewing size, with water staining, window frames and material transitions holding up at 100% zoom, although a few railing paths and small structural junctions were not physically convincing.

Typography and transparency held up

Four prompts explicitly required rendered text, and all four spelled the requested words correctly in this fixed-seed run. The author deliberately frames that as four observed samples rather than a text-accuracy claim. A bilingual storefront sign rendered both requested lines, in two languages, in the requested order, and an English neon sign followed the perspective of the storefront rather than sitting flat on the image.

Typography accuracy did not guarantee content accuracy, though: an educational infographic spelled every label correctly (“CLOUD”, “RAIN”, “SOIL”, “RIVER”) while its cloud panel already showed rain, blurring the distinction between the first two stages. Not everything obeyed the prompt either; a food editorial that asked for a top-down view came out closer to a 45-degree overhead angle.

The transparent sample carried a real alpha channel: of 1,048,576 pixels, 989,473 were not fully opaque. That verifies genuine transparency rather than a white or checkerboard background painted into the image, which matters for asset pipelines where many image generators fake cut-outs.

What the test does not tell you

The author is explicit that these figures describe one configuration, not a universal model benchmark. Batching, reduced-step schedules, quantization, CPU offload and alternative inference runtimes were all left untested. There is also a licensing constraint: Qwen-Image-2.1 is distributed under the Qwen Research License Agreement, which limits the model to non-commercial research or evaluation unless a separate commercial license is obtained.

Why it matters

For practitioners weighing a 96 GB workstation GPU, this turns local deployment into a cost calculation rather than guesswork. The model runs in full BF16 with peak allocations between roughly 37 and 57 GiB, leaving substantial headroom for batching or larger resolutions, and the ~384 images-per-GPU-hour figure at 1 MP is directly usable for capacity planning. The super-linear time cost of native 2K gives teams a concrete basis for choosing between generating at high resolution and generating at 1 MP before upscaling. The verified alpha channel addresses a common weakness in image generation for production asset work, and the typography results, while a small sample, suggest text rendering that holds up outside English. The non-commercial license remains the constraint no amount of GPU memory solves.