ViMax

Most AI video generation tools stop at the clip level — you get footage, not a story. ViMax addresses the layer above that: it wraps a Director, Screenwriter, Producer, and Video Generator into a single agentic pipeline that accepts a text idea or existing script and runs end-to-end through scene planning, dialogue writing, visual sequencing, and clip generation. The entry points are two scripts — `main_idea2video.py` for freeform prompts and `main_script2video.py` for structured inputs — with a web UI also available. Each step is handled by a dedicated agent; you configure once and the pipeline runs without manual handoffs between stages.

The differentiating capability is character and scene continuity across a full narrative arc. Where single-clip generators treat each generation as independent — producing consistency drift across scenes — ViMax’s Director and Producer agents carry context forward, so the character who appears in scene one is the same character in scene five. For educational content, children’s videos, and marketing storyboards where brand or character consistency is a hard requirement, that architectural choice solves a real production problem.

ViMax fits teams prototyping narrative video workflows, screenwriters validating scripts visually, and educators building explainer series — anywhere a multi-scene output is the goal and API costs per render are acceptable. It breaks down when production volume scales: every scene routes through Google Veo, Nanobana, and an LLM provider, so costs accumulate per clip rather than per seat. Teams running high-volume or commercial-scale pipelines will find the third-party API dependency becomes the ceiling. The project carries MIT licensing and the source is fully open, so teams with engineering capacity can swap in different backend providers, but that requires modifying the pipeline config and is not a zero-effort change.

Backend dependencies as described in the repository require API keys for Google Veo (video generation), Nanobana (image generation), and a compatible LLM provider for the script and direction agents. The repository documents a `configs` directory and pipeline architecture, with agent logic split across `agents`, `pipelines`, and `prompts` folders. A separate `Communication.md` describes inter-agent coordination. Self-hosting the framework is the default model — no vendor cloud is involved — but the underlying generation calls leave your infrastructure and reach third-party APIs on every run.