Pruna AI API

FLUX Fast is Pruna AI's speed-optimized version of Black Forest Labs' FLUX.1 [dev] image generation model. Pruna applies its own compression stack to the base model, combining flux caching (reusing outputs from earlier diffusion steps) with its diffusers2 compiler. On FLUX.1 [dev], Pruna reports this combination gives a 2.7x speedup over the original model, while keeping output quality close to the source on standard metrics such as LPIPS, SSIM, and PSNR. Pruna markets FLUX Fast as the fastest FLUX endpoint available, aimed at cases where turnaround time matters, such as interactive apps and high-throughput generation pipelines. It suits developers who want FLUX.1 [dev] quality without the latency of the standard model.

P-Image is Pruna AI's own text-to-image model, generating images in about one second at $0.005 each. Pruna describes it as building on prior open image models from Black Forest Labs (the FLUX series) and Alibaba (Qwen Image, Z-Image) rather than a single from-scratch architecture, though the company hasn't published full architectural details. Pruna positions P-Image as matching the quality of state-of-the-art image models while running about 30 times cheaper and 50 times faster, per its own figures. It's built for exact prompt adherence and reliable, controllable text rendering inside generated images. Use cases include real-time or high-volume applications, such as advertising and gaming assets, where both cost and latency matter.

P-Image LoRA is Pruna AI's inference endpoint for running custom-trained LoRA (Low-Rank Adaptation) adapters on top of P-Image. Rather than a fine-tuned model in itself, it's the serving side of a pair, users train a LoRA with Pruna's separate P-Image Trainer, then deploy it through P-Image LoRA to generate images with that custom style, subject, or character baked in. Trained adapters can be shared through Pruna's P-Image LoRA collection on Hugging Face. It carries the same fast, Pruna-optimized generation as the base P-Image model. It suits developers who want P-Image's speed and pricing combined with a personalized model, without retraining or hosting a full image model themselves.

Wan 2.2 Image is Pruna AI's optimized version of Alibaba's Wan 2.2 model running in image-generation mode, rather than the video mode Wan is best known for. Pruna applies its compression engine to the base model, generating 2-megapixel images in 3-4 seconds at $0.02 each. Pruna reports its version running 2.4 times faster than SeedDream, 1.8 times faster than FLUX1.1 [pro], and 1.1 times faster than its own earlier Wan 2.1 Image model. The model is tuned toward cinematic, aesthetically strong output rather than speed alone. It suits text-to-image workloads where a photographic, cinematic look matters, such as creative and marketing content, drawing on Wan's video-model lineage for composition.

Z-Image Turbo is Pruna AI's optimized version of Z-Image, a 6B-parameter open-source text-to-image model from Alibaba's Tongyi Lab, the team behind Qwen, released under Apache 2.0 in November 2025. The base model uses a Scalable Single-Stream DiT architecture and is distilled with Decoupled-DMD to just 8 sampling steps, which is what "Turbo" refers to. Pruna layers its own caching, compilation, and quantization on top for further speed, with generation reported at under one second on enterprise hardware. Z-Image ranks first among open-source image models on the Artificial Analysis leaderboard and supports native Chinese and English text rendering. Pricing is tiered by output resolution, from 0.25 cents at 0.5 megapixels up to 2 cents at 4 megapixels.

FLUX Kontext Fast is Pruna AI's speed-optimized endpoint for Black Forest Labs' FLUX.1 Kontext, an image editing model. It takes an input image and a text prompt describing the change, and returns the edited image. Pruna runs the model through its own optimization framework to reduce latency compared with its standard flux-kontext-dev endpoint, and describes the result as an ultra fast Kontext endpoint. Pruna has not published benchmark scores for it. Generation is billed at $0.01 per image. Parameters such as guidance strength, inference steps, and output quality can be adjusted to trade speed against fidelity. It suits developers who need instruction-based image edits in interactive apps or high-volume pipelines, where turnaround time and per-image cost matter.

P-Image Edit is Pruna AI's image editing model, the image-to-image counterpart to its P-Image text-to-image model. It takes a text instruction and one or more input images (up to four) and returns an edited image, typically in one to two seconds at $0.01 per image. It supports multi-image refinement, layout control, and style-preserving transformations, and is built to follow prompts closely. Pruna describes it as about 30 times faster and 3 times cheaper than comparable state-of-the-art editors, per its own figures. No independent benchmark scores were found. Typical uses include photo retouching, product visualization, and creative iterations such as changing backgrounds, adjusting lighting, or adding elements. It suits developers who need consistent edits within tight latency and cost budgets, such as production pipelines and high-volume applications.

P-Image Edit LoRA is Pruna AI's inference endpoint for running custom-trained LoRA (Low-Rank Adaptation) adapters on top of P-Image Edit, its image editing model. It is the serving side of a pair. Users train a LoRA with Pruna's separate P-Image Edit Trainer, then pass it to this endpoint to apply a specific edit style, such as a consistent transformation or look, to input images. A community collection of trained adapters is available on Hugging Face. The base P-Image Edit model is described as generating edits in about one to two seconds and accepting multiple input images, and this endpoint builds on it. Each output image costs $0.01. It suits developers who need repeatable, task-specific edits through an API without training or hosting a full image editing model themselves.

P-Image Ideogram is a text-to-image model from Pruna AI, built in collaboration with Ideogram, that generates an image from a text prompt in under a second. It inherits Ideogram's strength at rendering readable, correctly spelled text inside images. Prompts can be plain natural language or structured JSON with hex colors and bounding boxes for exact text placement. Output is available at 1K and 2K resolution, with eight aspect-ratio presets and custom width and height. A single thinking control with four levels (very low, low, medium, high) sets the trade-off between speed and quality from one endpoint. Generation starts at roughly 0.4 seconds at 1K, and each image costs $0.015. It suits developers building marketing graphics, ad creative, thumbnails, posters, and packaging mockups where headlines and labels must be legible.