Benchmarks

The Flux.1 AI Image Generation Demo was developed in partnership with NVIDIA for CES 2025, and currently supports only NVIDIA GPUs.

FLUX.1 is an AI model developed by Black Forest Labs for high-quality text-to-image generation. This demo showcases the performance benefits of running the FLUX.1 model with FP4 and FP8 precisions with FP16, and generated images can be visually compared after the demo has been run.  

Model

This demo can use either the Flux.1-Schnell model or the FLUX.1 [dev]  model. 

Licenses for the Procyon AI Inference Benchmarks do not include a commercial license for the FLUX.1 [dev] model. To use the FLUX.1 [dev] model with Procyon, you need an appropriate license for FLUX.1 [dev] from Black forest labs.

You can add a Hugging Facetoken with read rights to the FLUX.1 [dev] model in the Procyon settings to enable automatic downloading using Flux.1 [dev], or you can download and add the model manually. 

Installing Models Manually

Download all the model files and place into the following directories. 

Flux.1-Schnell

This can be downloaded from https://huggingface.co/Futuremark/FLUX.1-schnell-onnx/tree/main 

C:\PROGRAMDATA\UL\PROCYON\CHOPS\DLC\AI-IMAGEGENERATION-BENCHMARK\MODELS\ONNX_CUDA_OPTIMIZED\Futuremark\FLUX.1-schnell-onnx

C:\PROGRAMDATA\UL\PROCYON\CHOPS\DLC\AI-IMAGEGENERATION-BENCHMARK\MODELS\ONNX_CUDA_OPTIMIZED\black-forest-labs\FLUX.1-dev-onnx\

Demo Settings

FLUX.1 Demo System Requirements

Storage: 

• 50GB free disk space

FP16 precision:

• NVIDIA GPU with 24GB VRAM (35GB to fully load pipeline)

FP8 precision:

• NVIDIA GPU with 12GB VRAM (23GB to fully load pipeline)

FP4 precision:

• NVIDIA GPU with 8GB VRAM (15GB to fully load pipeline)

If enough free VRAM is detected the benchmark will load the pipeline fully into video memory. This speeds up the runtime of the demo.  

This can be forced on or off via the offloading command line parameter in the workload definition file.