Pruna cuts FLUX.2 [flex] latency 3.7x
Pruna AI and Black Forest Labs have shipped an optimized FLUX.2 [flex] on BFL's official API, cutting single-H200 generation time from about 48 seconds to 13 seconds per image.
Inference-optimization startup Pruna AI announced that it worked with Black Forest Labs (BFL) to speed up FLUX.2 [flex], BFL's image model built for text rendering and typography. The accelerated version is now served through BFL's official API at the flux-2-flex endpoint.
On a single H200, Pruna says text-to-image generation at 1024x1024 fell from 48.53 seconds to 13.11 seconds per image, a 3.7x speedup. Image editing fell from 104.79 seconds to 27.63 seconds, or 3.85x. With four GPUs, optimized generation took 4.98 seconds and editing 9.94 seconds.
Pruna says the optimized model keeps the original's features: guidance-scale control, accurate text rendering, and support for up to eight reference images for multi-image editing. In Pruna's own comparison, the fast version is listed at $0.05 per image, against $0.06 for the base model. It scored 0.8806 on LongTextBench (base: 0.8681) and 0.7666 on the OneIG text-rendering subset (base: 0.7917). GPT Image 1.5 (high) scored higher on LongTextBench at 0.9152, and FLUX.2 [max] scored higher on OneIG at 0.8230. Pruna says the optimized model sits on the speed-quality Pareto front and calls it the cheapest high-quality option on the market.
All figures are self-reported by Pruna. The post does not say which optimization techniques were used, and it does not say whether the speedups carry over to BFL's production serving hardware.