LibTV Conversational Image and Video Generation

Problem

In AI image and video platforms such as LibTV, a single user request may require multiple backend operations: uploading reference assets, choosing models, splitting a short drama into shots, applying style edits, and collecting generated media. If a client-side agent expands prompts, invents style descriptors, or manually orchestrates model calls, it can disrupt the backend workflow and make integration brittle. This skill exposes a session-based interface so the user can send natural-language creative instructions while the backend agent handles understanding, orchestration, and model invocation.

How It Works

• Create session / send message: Use create_session to start a session or send an instruction such as “generate an anime video” or “replace the paper boat with a heart.”
• Upload and edit: When the user supplies a local image or video, call upload_file to obtain an OSS URL, then submit the original instruction together with that URL. Editing tasks depend on the uploaded asset reference.
• Poll and retrieve: Query progress with query_session, using sessionId and incremental --after-seq values, typically every 8 seconds. Treat an assistant message containing a result URL as completion, then use download_results to download images or videos in bulk.
• Project and display: create_session may return projectUrl; after the task finishes, show both the result link and the project canvas link. Do not expose the canvas link during in-progress polling.

Boundaries and Notes

This skill fits environments where LibTV credentials and session context are available. It supports text-to-image, text-to-video, image-to-video, video continuation, local edits, element replacement, style transfer, and longer productions such as short dramas, MVs, or promotional videos. Uploaded files are limited to images and videos, usually under 200MB. Complex tasks may take time, so use the polling timeout and retry behavior. The client-side agent should remain a messenger: upload assets, pass the user's request through unchanged, and deliver the final result and project canvas link.