Pruna AI launches faster FLUX.1 endpoint on Replicate
Pruna AI has released FLUX-juiced, an optimized FLUX.1 endpoint on Replicate that it says runs up to 2.6x faster than the official API without losing image quality, along with a benchmark comparing inference providers.
Pruna AI announced FLUX-juiced, an optimized version of Black Forest Labs' FLUX.1 image model, now live on Replicate. The company says it delivers up to 2.6x faster inference than the official Replicate API "without sacrificing image quality." It notes that FLUX.1 often takes more than 6 seconds per image, even on H100 GPUs.
Pruna says the speedup comes from combining compiler-level graph optimization with selective caching of heavy operations such as attention layers, so inference can skip redundant computation. The company did not detail the internals. It says the techniques are general and plug-and-play through its paid Pruna Pro pipeline, and can apply to nearly any diffusion-based image model. A free open-source version is also available.
Alongside the endpoint, Pruna published InferBench, a Hugging Face Space comparing FLUX-juiced with the FLUX.1 [dev] endpoints from Replicate, Fal, Fireworks AI and Together AI. Pruna notes that these providers don't always disclose the optimizations they use. All endpoints were tested with 28 inference steps, 1024×1024 resolution and a guidance scale of 3.5. Of the providers, only Replicate reported its hardware, an 80GB H100.
The prompts came from DrawBench, HPSv2, GenAI-Bench, GenEval and PartiPrompts. Outputs were scored with ImageReward, VQA-Score, CLIP-IQA, CLIP-Score, CMMD, ARNIQA and a Laplacian-variance sharpness measure.
The available excerpt stops before the per-provider results, so speed, cost and quality figures for each competitor are not covered here. Pruna points readers to the InferBench Space for the full results. Pruna both built FLUX-juiced and ran the benchmark, so the results have not been independently verified.