Flux API (Black Forest Labs)

Generates photorealistic images from natural language prompts using a selection of Flux model variants (Pro, Dev, Schnell, or FLUX.2 family) optimized for different speed/quality tradeoffs. The API accepts text prompts and routes them through the selected model's inference pipeline, which applies diffusion-based generation with architectural optimizations for prompt adherence and visual fidelity. Users select model variant at request time, enabling dynamic quality/latency tuning without redeployment.

Enables guided image generation by conditioning on multiple reference images (up to 10) alongside text prompts. The API accepts reference images and applies them as control signals during the diffusion process, allowing style transfer, object replacement, pattern matching, and composition guidance. Implementation uses multi-image conditioning architecture where reference images are encoded and injected into the generation pipeline to steer output toward desired visual characteristics while respecting the text prompt.

Allows per-request specification of output image dimensions (width and height in pixels) up to a maximum resolution determined by model variant. The API accepts width and height parameters in the request payload and generates images at the specified dimensions. FLUX.2 [max] supports up to 4MP output; other variants have lower maximum resolutions (unspecified). Implementation likely uses adaptive inference scaling or resolution-aware model conditioning to generate at arbitrary dimensions within the supported range.

Exposes multiple Flux model variants (Pro, Dev, Schnell, FLUX.2 [klein/pro/flex/max]) with documented or claimed performance characteristics, allowing developers to select the optimal variant per request based on latency and quality requirements. FLUX.2 [klein] is positioned as 'fastest image model to date' with sub-second inference; FLUX.2 [max] targets production-grade 4MP photorealistic output. Implementation routes requests to the selected model's inference endpoint, with no automatic fallback or variant selection logic — developers must explicitly choose.

Supports generation of multiple images in sequence or batch through repeated API calls, with pricing that scales based on output dimensions and number of reference images used. The pricing calculator interface shows width, height, and reference image count as parameters, suggesting per-request pricing is computed as a function of these variables. No documentation of batch endpoint, async job submission, or bulk discounts — pricing appears to be per-request with no volume optimization.

Offers free trial access to Flux models with the messaging 'Try FLUX.2 for free' on the website, but specific trial limits, credit allocation, duration, and model variant availability are not documented. Implementation likely uses a credit-based system where free tier users receive an initial credit allocation that depletes with each request; exact credit values and replenishment policies are unknown. No documentation of free tier restrictions (e.g., lower resolution, longer latency, or limited model variants).

Flux models are available through third-party API providers (Replicate, Together AI, fal.ai) in addition to direct Black Forest Labs API access. These providers offer standardized API interfaces, SDKs, and integration tools that abstract away direct Flux API complexity. Implementation routes requests through the chosen provider's infrastructure, which handles authentication, rate limiting, billing, and request routing to Flux inference endpoints. Developers can choose providers based on preferred SDK language, pricing, or existing integrations.

FLUX.2 [klein] is a lightweight model variant optimized for sub-second inference latency on capable hardware, enabling real-time or near-real-time image generation in interactive applications. Implementation uses architectural optimizations (likely reduced model size, quantization, or inference acceleration) to achieve sub-second generation time. Positioning emphasizes speed over maximum quality, making it suitable for latency-sensitive use cases where instant feedback is critical.

FLUX.2 [max] is a production-grade model variant optimized for maximum output quality and resolution, supporting up to 4MP (megapixel) photorealistic image generation. Implementation prioritizes visual fidelity and detail over inference speed, using full-capacity model architecture and inference optimizations for quality. Positioning targets professional use cases (product photography, marketing, design) where image quality directly impacts business outcomes.

Flux models are positioned as having strong 'prompt adherence,' meaning they accurately interpret and render text prompts into visuals that closely match the described intent. Implementation uses training techniques (likely RLHF, instruction tuning, or similar) to align model outputs with user intent as expressed in natural language. This is a qualitative capability rather than a quantifiable metric, but it's emphasized as a key differentiator in marketing materials.