fal Worlds: AI-generated worlds that you can explore and direct in real time

fal launches Worlds, a platform to explore and direct AI-generated worlds in real time, featuring LingBot World, H3 Max Director, and Abot World.

Until now, AI video generation has largely involved submitting a prompt, waiting for the result, and repeating the process to create another sequence. With fal Worlds, fal introduces a different approach: continuously generated environments that users can navigate, influence, and direct as scenes unfold. The platform brings together interactive world models including LingBot World 2, H3 Max Director, and Abot World. Underneath is the World Model Accelerator, a new inference infrastructure designed to reduce the delay between user actions and model responses. From Generated Video to Interactive Environments

Traditional video generation models typically produce a sequence from text, an image, or multiple references.

The resulting content has a duration and structure largely determined at generation time.

World models explore a different approach. They attempt to represent how an environment evolves based on its current state and, in some cases, actions performed by a user.

Users can potentially move a camera, change direction, or instruct a scene to evolve while its visual content continues being generated.

With Worlds, fal brings these capabilities into a single interface for experimenting with different models.

The platform highlights three modes: Walkable, for moving through an environment; Directable, for controlling a scene during generation; and Explorable, for navigating an interactive world.

These categories represent distinct experiences. Not every model supports the same interactions, and directing a video sequence does not necessarily mean the system maintains a complete, persistent simulation of its environment. Multiple Models Available Through One Platform

The fal Worlds catalog includes models designed for both navigation and scene direction.

H3 Max Director generates a continuous video stream that can be modified through prompts sent during an active session. Users establish an initial situation and then change how it develops without restarting generation for every instruction.

H3 Max Reference Director adds reference images to guide the appearance of characters, objects, and environments across a sequence.

LingBot World 2, provided by Robbyant, allows users to explore interactive environments through movement and camera controls.

Abot World, associated with Alibaba, similarly offers navigable environments with movement and viewpoint controls.

The selection illustrates two broad approaches: directing a continuously generated audiovisual stream and exploring an environment that responds to navigation commands.

Each model has its own interface, capabilities, and usage conditions. H3 Max Director: Changing a Scene While It Plays

H3 Max Director illustrates the continuous direction approach.

Instead of waiting for a completed generation before accepting a new request, it maintains an active session with the user.

The initial prompt establishes the setting, characters, situation, or narrative direction.

The model then begins generating a video sequence with audio.

Additional prompts can be submitted during playback to modify what happens next.

For example, a user could ask a character to leave a room, approach another person, or continue an action in a different location.

The model uses information from earlier segments to maintain some degree of visual and narrative continuity.

New instructions affect segments that have not yet been generated. Previously delivered footage is not rewritten.

According to the H3 Max Director documentation, sessions use WebRTC video and audio streams with configurable resolution and aspect ratio.

The approach is relevant to both narrative experimentation and applications requiring audiovisual responses during an interaction. Reference Images for Character Consistency

H3 Max Reference Director extends this workflow by allowing users to provide visual references.

These images can define the appearance of characters, clothing, objects, or settings.

Text instructions can still be submitted during the session to guide what happens next.

The system attempts to evolve the action while preserving the visual elements established through the references.

Its technical interface includes settings for reference fidelity, retained context, and how newly submitted instructions affect the stream.

These controls are particularly relevant to workflows involving recurring characters or a consistent visual style across several scenes.

However, references do not guarantee perfectly stable identities under every condition. Character and setting consistency still depend on scene complexity and duration. LingBot World 2 and Abot World: Exploration Instead of Playback

With LingBot World 2 and Abot World, interaction places greater emphasis on navigation.

Users can issue movement and camera commands to select a direction and observe how the model changes the environment.

This more closely resembles navigating a virtual space than producing a linear film.

An important difference remains between these systems and conventional game engines.

In a traditional game, geometry, textures, objects, and many environmental rules are explicitly represented within a graphics engine.

A generative world model may instead produce the next visual observations based on context and received actions.

This creates possibilities for interactive prototyping, but it also introduces stability challenges.

Object persistence, spatial consistency, memory of earlier actions, and reliable physical interactions remain active research problems.

The ability to navigate a generated environment therefore does not imply that the model has every property of a conventional three-dimensional simulation. World Model Accelerator: Reducing Interaction Delays

To support these experiences, fal introduces the World Model Accelerator, or WMA.

It is an inference infrastructure designed specifically for models producing interactive content and continuous streams.

In conventional video generation, a request may be submitted, placed in a queue, processed, and