Video Pipeline

Validates inputs, reserves credits, and creates the job record.

The system checks your product image, validates all parameters, estimates credit cost, and reserves credits from your balance. If validation fails, no credits are consumed.

AI agents generate the video script and plan keyframe prompts.

A specialized AI agent analyzes your product image, persona, and settings to write a natural UGC-style script. It plans the visual keyframes — including camera angles, expressions, and product interaction shots.

Generates the anchor keyframe — the visual foundation for the entire video.

The anchor keyframe establishes the persona’s appearance, lighting, and environment. All subsequent keyframes reference this image to maintain visual consistency throughout the video.

Generates all video segments from keyframe pairs.

Each segment is created from a keyframe pair using Veo 3.1 or Kling 3.0. Segments generate with streaming overlap — after 2 consecutive keyframes pass QA, video generation begins in parallel with remaining keyframes.

Stitches all video segments into a single continuous video.

The FFmpeg microservice concatenates segments with smart trimming, handling frame alignment and transition smoothing between segments.

Applies voice-over and mixes background music.

The script is performed using ElevenLabs STS (Speech-to-Speech) with the selected voice. The audio is synced to the video timeline and mixed with optional background music from the library.

Generates word-level synchronized captions.

Captions are generated with karaoke-style word highlighting using ASS subtitle format. Supports 29 languages with automatic language detection and RTL layout for Arabic and Hebrew.

Final render, upload, and delivery.

The FFmpeg microservice performs the final render (burning in captions, mixing audio tracks), uploads the finished video to secure storage, finalizes credits (reserved → consumed), and triggers webhook notifications. The video is ready for download or social auto-posting.