Generative image and video models

fal.ai hosts a range of generative media models a developer can call to produce an image or a video clip from a prompt rather than building that capability in house. A Beam agent submits a prompt when a workflow step calls for a visual asset, such as a placeholder image for a draft page, and reads back the result fal.ai returns. It attaches that asset to the record it is working on. Outputs that look off brand, low quality, or unsuitable for a customer facing use are set aside for a person to review before publishing anything with them.

Developer facing generation API

fal.ai is built as an API first platform, so a developer calls it directly from code rather than through a separate app, which is what lets a Beam agent trigger a generation step as part of a larger workflow. The agent sends the prompt and any parameters a task calls for, reads the response, and moves the result into whatever record or asset library the workflow uses next. API errors, timeouts, or a result the agent cannot match to the request that triggered it are logged and handed to a person rather than retried indefinitely without oversight.

Prompt driven asset creation

Because fal.ai's models take a text prompt as their main input, the quality of what comes back depends heavily on how that prompt is written, which is why a Beam agent works from prompt templates a team has already approved rather than writing prompts from scratch. It fills in the details a specific task needs, such as a product name or a described scene, and submits the completed prompt. Requests that fall outside the approved templates, or that a workflow marks as sensitive, such as anything involving a real person's likeness, are routed to a person instead of generated automatically.