The UL Procyon AI Image Generation Benchmark generates a set of images (default 16), separated into batches. These batch sizes vary per image generation model, with SD1.5 having a batch size of 4, and SDXL having a batch size of 1.
On completion of the benchmark, the overall score is calculated by taking into account the average inference times of the models and a scaling constant used to bring the score in line with the traditional range for UL benchmarks, depending on the Stable Diffusion model used. The higher the score, the better the performance.
Stable Diffusion 1.5 AI Image Generation score = K * (20/run time) Where K = 5000
Stable Diffusion XL AI Image Generation Score = K * (120/run time) Where K = 5000
K is a scaling constant used to bring the score in line with the traditional range for UL Benchmarks.
In addition to the main benchmark score, the results screen also displays the following sub scores:
- Overall time taken in seconds
- Overall image generation speed as seconds per image.
The results screen displays further details for each batch of generated images, as well as the final generated images. While the images should be similar between each run as they use the same seed, they can vary slightly due to the nature of adding random noise. This has no meaningful effect on benchmark scores.
- Text encoder: converting text prompt into tokenized text.
- UNET: Takes tokenized text, adds random noise, then loops denoising steps to create an image in the latent space.
- VAE: decodes latent image into final (actual) image output.
- Pipeline: all the above steps