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

Description and use

On an RTX PRO 6000, Qwen‑Image‑2.1 produces 14 1‑MP images in ~9.4 s each and a 2048×2048 image in 52 s, using 36.8 GiB and 56.5 GiB peak memory respectively, achieving ~384 images per GPU‑hour.

Problem

Researchers need to know the image quality and inference speed of Qwen-Image-2.1 on a 96 GB workstation GPU.

Outcome

On a RTX PRO 6000, Qwen‑Image‑2.1 generates 14 1‑MP images in ~9.4 s each and a 2048×2048 image in 52 s, using 36.8 GiB and 56.5 GiB peak memory respectively, achieving ~384 images per GPU‑hour.

Workflow

• 1. Install PyTorch 2.11.0+cu128 and Diffusers.
• 2. Load QwenImage21Pipeline with BF16 precision, 40 steps, batch size 1.
• 3. Run 15 fixed‑seed prompts covering typography, photography, architecture, illustration, and a transparent asset.
• 4. Record per‑image generation time and peak GPU memory via PyTorch.
• 5. Aggregate results and calculate average speed and memory usage.

Evidence

On September 24, 2026, we ran 15 fixed‑seed prompts through the official Diffusers pipeline.

Source

BF16, 40 steps, batch size 1

Source

9.378 seconds per image on average

Source

56.516 GiB peak allocated GPU memory

Source