Krea turns AI video prompts into live controls

Krea has launched Realtime Director, an interface for steering a continuous AI-generated video stream while it plays. Users can change the scene by submitting new prompts during a session, while an optional reference image controls the opening frame.

Most AI video APIs follow a one-prompt, one-file workflow. Realtime Director gives developers a long-lived, stateful session instead, with ongoing control messages, WebRTC media delivery, and time-based billing. The underlying model is fal’s H3 Max Director, which was built for continuous generation.

Prompts become live controls

Krea’s current interface keeps the workflow compact:

• Starting frame: An optional reference image defines how the stream begins.
• Live direction: Text prompts are the only mid-stream steering input.
• Prompt changes: Each new instruction becomes the active direction while the video continues.
• Storage: Completed videos save automatically to Krea Assets.

The continuous session allows characters, settings, and visual context to carry across prompt changes. Clip-based systems often generate separate segments and splice them together, creating visible cuts or identity drift at each boundary.

Two minutes of rolling context

H3 Max Director is an autoregressive video model with up to two minutes of contextual memory. In plain English, it generates each new section from the recent stream and latest instruction, preserving a rolling history of the characters, setting, audio, and action.

A session can run for up to 15 minutes by default, while the model’s stated context window covers up to two minutes. Developers should treat those as separate limits: a longer session does not retain every earlier frame in active context. Applications that depend on persistent identities or locations should test for drift as earlier material leaves the rolling window.

WebRTC carries the session

Fal delivers the generated stream over WebRTC at 24 frames per second with 32 kHz audio. Output is generated in 10-second chunks, with standard modes covering 480p or 768p and aspect ratios of 16:9, 9:16, or 1:1.

Each live instruction includes a strictly increasing prompt_version. Clients should serialize updates, increment that value for every prompt, and avoid reusing version numbers when users submit directions quickly. The announcement does not specify prompt-to-frame latency, so applications with interactive controls should measure response time under realistic network conditions.

A direct integration through the fal API therefore needs to manage several pieces of session state:

1. Create and maintain the long-lived generation session.
2. Receive video and audio through WebRTC.
3. Send ordered prompt updates with increasing version numbers.
4. Handle connection loss, reconnection, and media capture.
5. Track elapsed generation time for cost controls.

The meter runs by the second

Krea provides the hosted interface, while fal exposes the model for application development. Fal’s published API terms are:

At standard resolution, the promotional rate equals $1.20 per minute, and the list rate equals $4.80 per minute. A full 15-minute session would cost $18 during the promotion or $72 at list price. The 1080p tier doubles those totals.

Fal’s rates apply to direct API use. Krea’s launch announcement does not publish a separate price for sessions created through its interface.

Applications that need the next frame now

Continuous generation supports applications where the next scene depends on live input:

• Audience-directed stories that change while they stream.
• Interactive backgrounds for games, broadcasts, and performances.
• Persistent AI channels with recurring characters and locations.
• Previsualization sessions where a director revises a scene verbally.
• Generative environments or background footage produced on demand.

Current constraints narrow the production range. Standard output tops out at 768p, with 1080p priced at twice the rate. Krea offers no mid-stream image conditioning, reference-video input, or dedicated camera controls. Text remains the only live steering channel after the initial frame, and sessions beyond 15 minutes require approval.

Quality claims need runtime tests

Fal describes Director as the streaming variant of its H3 Max family. The company says the base model holds top-three positions across Artificial Analysis leaderboards for video editing, text-to-video, and image-to-video with audio, and can render a five-second clip in under three seconds.

Director’s prompt latency, long-session identity retention, audio continuity, and WebRTC stability require separate measurements because those rankings cover base-model output quality. The streaming layer introduces network behavior, ordered control messages, rolling context, and extended generation time that clip benchmarks do not capture.

Tests to run before shipping

Teams evaluating Realtime Director should test the complete session under expected production conditions:

• Measure the delay between a prompt update and a visible change.
• Check whether 10-second generation boundaries affect responsiveness.
• Track character and setting consistency beyond the two-minute context window.
• Simulate dropped WebRTC connections and verify prompt-version recovery.
• Confirm recording, export, and storage behavior for direct API sessions.
• Set spending limits for maximum duration and resolution.

Realtime Director gives developers a stateful, steerable media stream that can react throughout a session. That model supports interactive video products without requiring an application to generate, download, and splice a new clip after every user instruction.