AI filmmaking with an infinite canvas and a memory for faces

What it does

AICON is a full-stack workbench for AI-driven video creation built on FastAPI, Vue 3, PostgreSQL, Redis, and Celery. It ingests long-form text, extracts characters and scenes, and walks through storyboards, keyframes, transitions, and final video synthesis. The pipeline ends with automated distribution—currently centered on Bilibili—including upload, titles, and tags. The goal is to keep the whole production chain inside one system instead of bouncing between a dozen browser tabs.

The interesting bit

The core is an infinite canvas where text, image, and video nodes are wired together by hand, and a side-panel Canvas Assistant can auto-build a production chain from a single creative prompt. The assistant sets up the nodes and connections but deliberately stops short of firing generation jobs, forcing the human to review before burning API credits. Character reference images are automatically injected into downstream keyframe nodes to reduce the usual “who is this person now” problem across cuts.

Key highlights

• Natural-language-driven infinite canvas with node dependencies, reference-image overlays, and generation-history rollback
• Movie Studio pipeline: script → characters → scenes → storyboards → keyframes → transitions → final video with BGM and sound effects
• Picture Narration mode for batch-generating illustrated short videos with auto-generated voice and subtitles
• Character-consistency guardrails via reference-image injection across keyframes
• Async Celery-backed architecture for long generation chains; pluggable model providers via custom Base URL

Caveats

• Distribution currently centers on Bilibili; other platforms aren’t named in the sources.
• The default model gateway (api.aiconapi.me) is the author’s self-hosted relay, though the code allows swapping to any OpenAI-compatible endpoint.
• The Canvas Assistant creates the workflow skeleton, but actual generation tasks must be triggered manually.

Verdict

Worth a look if you’re producing serialized AI video or illustrated shorts and want a structured alternative to prompt-hopping in chat interfaces. If you just need a one-off text-to-video clip, the full stack is probably overkill.