fal is a generative media inference platform made by the company Features and Labels. Inference means running a trained AI model to produce an output; generative media models produce images, video, audio, speech and 3D objects from a text prompt or an input image.
fal hosts more than 1,000 models from many labs, such as Black Forest Labs, Google, OpenAI and ByteDance, behind one API (an interface that a program calls over the internet). A developer picks a model, gets an API key and sends a request, for example a prompt, with the Python or JavaScript client or with plain HTTP; the result comes back with a link to the generated file. Calls can be synchronous or go through a queue, and some models stream results in real time. Teams can also deploy their own models on fal Serverless, GPUs that scale with traffic, or rent dedicated GPU machines (fal Compute) for training.
fal is a hosted service with usage-based pricing: Model APIs are billed per output, such as per image or per second of video, and GPUs per hour.