Pruna AI launches Flux benchmarking playground
Pruna AI has released a public test app that compares Black Forest Labs' Flux image models with its own compressed versions, after a rise in customer questions about running Flux on cheaper GPUs.
Model-compression startup Pruna AI wrote in a technical blog post that it now gets three to five requests a week from companies about its public quantized versions of Flux. The requests started after users found that the compressed models fit on an Nvidia A10G GPU. Companies wanted to know how much faster the models are, whether quality holds across image sizes and batch sizes, and which hardware they support.
To answer those questions, Pruna built a "Flux Playground." The company calls it a test URL with a simple design. The app compares the FLUX.1 [schnell] and FLUX.1 [dev] base models with Pruna's optimized "Turbo" and "Fast" versions across more than 60 prompts. It reports speed-up and cost figures and shows the images side by side. Because judging image quality is subjective, Pruna also lets customers collect live comparison votes from a panel of human raters. The post does not include the benchmark numbers.
The post describes Flux as a 12-billion-parameter "rectified flow transformer." Flux is based on flow matching, which moves from noise to an image along a straighter path than standard diffusion, so it needs fewer sampling steps. The schnell variant was trained with latent adversarial diffusion distillation and can produce images in one to four steps. The author notes that researchers still disagree about whether flow matching beats diffusion.
Pruna says its getting-started example takes four steps and 28 lines of code and needs an access token, which users can now get by email. The company also points to a customer case study in which Each AI reached production within days and cut cost and latency by about three times. That figure comes from Pruna and has not been independently checked.