LibTV’s Video Agent Takes AI Video Creation to the Next Level

Video Agent workflow in LibTV

Video creation typically requires multiple iterations: converting a story to a script, rearranging shots, maintaining character consistency, and repeatedly refining visuals. LibTV’s Agent automates the initial stages—generating a script, storyboard, shot arrangement, and the full production pipeline—then allows users to switch between a Storyboard view for rapid layout and a Node view for detailed adjustment of shots, assets, model parameters, and workflow steps.

Skill 1 – Moving a Figurine

Input: a static figurine image.

Upload the image.

Agent generates a three‑view preview (initial layout, shot selection, and final rendering).

User confirms the direction and proceeds to full video generation.

The Agent animates the figurine, turning a non‑moving object into an on‑screen character.

During generation the interface presents selectable options for each step (see image below).

Multi‑round confirmation screens allow the user to adjust formats or parameters before final rendering.

If a generated shot contains an artifact (e.g., a double‑face close‑up), the user identifies the problematic storyboard panel, instructs the Agent to regenerate that specific shot, and the issue is resolved without restarting the entire project.

Skill 2 – Luxury‑Feel TVC

Input: a single product image.

Select the “high‑end TVC” Skill.

Upload the product image.

Agent produces a three‑view preview.

User confirms the advertising direction.

Full video generation begins.

The three‑view preview is displayed (see image).

During generation a hallucination appears (a close‑up with two faces). The user locates the offending storyboard panel and commands the Agent to regenerate that shot. The regenerated version resolves the artifact.

Custom Skill – Stick‑Figure Performance

Goal: generate a stick‑figure avatar from an uploaded image and produce a performance video.

Upload the source image.

Agent creates a stick‑figure representation.

Three‑view preview is shown.

User confirms and renders the final video.

The resulting video matches expectations, demonstrating that users can author their own Skills and follow the same pipeline.

Key technical observations

All‑purpose Video Agent : Accepts multimodal inputs (text, image, video, audio, documents) and automatically decomposes them into scripts, shots, visuals, sound, subtitles, and edit tasks, producing director‑level output without manual re‑assembly.

Dual view switching : Storyboard view enables rapid creative layout; Node view provides fine‑grained control over any element in the pipeline, with real‑time linkage between the Agent and professional node‑based tools.

Iterative refinement : After initial generation, the Agent can re‑generate specific clips, edit subtitles, trim segments, reorder shots, add transitions, and optimize audio, all without restarting the project.

Extensible Skill library : Hundreds of pre‑built Skills (e.g., directing, self‑media, advertising, music video, short drama) can be invoked by natural‑language prompts; users can also author custom Skills to encapsulate personal prompts, asset conventions, and workflow steps.