Pruna AI claims 5x faster Flux-Kontext
Pruna AI says its compressed versions of the newly open-sourced Flux-Kontext image editor run up to 4.9x faster on one H100, though it offers only visual comparisons to support its claim that quality is unchanged.
Model-compression startup Pruna AI says it has made the open-source Flux-Kontext image-to-image model up to about five times faster. According to the company, the base model takes roughly 14.4 seconds to generate one image on a single H100.
Pruna built the "juiced" variants with its commercial Pruna Pro toolkit. It combined caching, factorization, quantization and compilation, and says compressing, deploying and evaluating the model took less than four hours. The company released three configurations and reports these timings for a 1-megapixel image over 30 steps:
- Lightly-Juiced: 4.5s (3.2x)
- Juiced: 3.7s (3.9x)
- Ultra-Juiced: 2.9s (4.9x)
Pruna says the speedups cut user wait times, serving costs and energy use. On quality, the company says the outputs stay "unchanged" and points to side-by-side sample images. The post gives no quantitative quality metrics and no independent benchmarks, and all timings come from Pruna itself.
The compressed model is available through a Pruna endpoint on Replicate. Pruna says users will be able to reproduce the compression with a configuration on the upcoming Flux-Kontext integration in Hugging Face. The post does not say how many parameters the compression changes or how the three variants differ in quality. It also doesn't say whether the configuration works outside Pruna Pro, which is a paid product.