An agent that runs your whole AI art studio, canvas included
Novanova Studio wires an AgentScope agent into an infinite canvas so image and video generation, editing, and history live in one continuous conversation.

What it does
It’s a self-hostable, Java/Spring Boot creation workbench for AI images and video. The centerpiece is an infinite canvas where text, reference material, images, videos, and generation results coexist, plus an agent that reads canvas state, picks tools, and kicks off tasks on your behalf. Underneath sit user accounts, credits, an admin backend, and adapters for OpenAI-compatible, Gemini, Anthropic, Seedance, MiniMax, Evolink, and custom template-based model endpoints.
The interesting bit
The agent isn’t a bolted-on chat box: frontend canvas tools report their real execution results back to the server-side agent, so the loop of “converse → tool → result → converse” actually holds. Long jobs go through Redis Stream consumer groups with task locks, failure recovery, and SSE push-back — the boring plumbing that makes video generation survivable.
Key highlights
- Agent profiles per scene (image, video, canvas) with tool calls for generation, editing, history, and node manipulation
- Workflows like
first-last-frame-to-videochain requirement clarification, prompt confirmation, frame generation, and synthesis - Per-model billing rules, including resolution/duration tiers for video; disabled models vanish from user dropdowns
- Object storage via Tencent COS, Aliyun OSS, or Qiniu Kodo; keys stay server-side
- Skills are configurable prompt-driven guided interactions, extensible without code
Caveats
- The Agnes channel’s chat adapter exists but isn’t wired into the main agent’s model factory yet
- README leans heavily on Chinese; the English translation exists but the docs and community are China-centric, and the README doubles as an ad board for API resellers
- “Novel-to-short-drama” and 3D director features are advertised in the tagline but not detailed in the README
Verdict
Worth a look if you want a self-hosted, credit-metered AI creation platform with a genuinely agent-driven canvas rather than a pile of disconnected generation forms. Skip it if you wanted a lightweight single-purpose tool — this is a full platform with PostgreSQL, Redis, and an admin console to feed.
Frequently asked
- What is Swayingleaves/novanova-studio?
- Novanova Studio wires an AgentScope agent into an infinite canvas so image and video generation, editing, and history live in one continuous conversation.
- Is novanova-studio open source?
- Yes — Swayingleaves/novanova-studio is open source, released under the GPL-3.0 license.
- What language is novanova-studio written in?
- Swayingleaves/novanova-studio is primarily written in Java.
- How popular is novanova-studio?
- Swayingleaves/novanova-studio has 501 stars on GitHub.
- Where can I find novanova-studio?
- Swayingleaves/novanova-studio is on GitHub at https://github.com/Swayingleaves/novanova-studio.