Repository navigation
Getting Started.en
中文 | English
🌐 This is a translation. The Chinese version is authoritative — if anything here conflicts with the Chinese original, the Chinese version prevails, and translations may lag behind it.
This page gets you up and running in the shortest possible time. Creators can start working after finishing this page; developers should continue with the Workflow Development Overview afterwards.
Unigen is an Electron desktop application supporting Windows / Linux / macOS.
- Recommended: download the portable build for your platform from Releases, unzip, and run.
- Build from source: see Local Setup.
| Strategy | Local load | Reference configuration | Best for |
|---|---|---|---|
| External model APIs only (built-in models disabled) | Only SQLite + LanceDB, modest data volume | 4-core CPU / 8 GB RAM / 2 GB disk, no GPU | Creators (recommended) |
| Built-in models enabled (local embedding / reranking) | Depends on the chosen models | Available models depend on your hardware | Users with hardware who want to save API costs or work offline |
When enabling built-in models, we recommend llmfit to assess what model sizes your machine can run. Note: llmfit does not directly recommend embedding/reranking models — estimate using its VRAM/RAM conversion method.
After first launch, click Settings (bottom-left) and configure your models:
| Model | Required? |
|---|---|
| Text model | Required |
| Embedding model | Required |
| Reranking model | Recommended |
These settings only need to be configured once and apply to all project types.
Creating a project requires choosing a project type:
- Script-to-Video — the only currently runnable project type, good for experiencing the full pipeline.
- Workflow Studio — for developers: the project type used to build workflows and export them as new project types (i.e., "the project that generates other project types"). Workflow-development / planning capabilities are built in (see Developer Guide).
- Other project types exist as placeholders for now — no deliverable workflow yet (see Roadmap).
About Script-to-Video: its code was generated by AI and only verified to be runnable — it has not been deeply polished. Since video-generation APIs are costly, we recommend first setting the execution target to stop at "prompt generation": inspect the generated prompts and test on a small scale before running the full video pipeline. If you spot quality issues, please report exactly what is wrong to help us improve the video Agent.
Security note: the current version trusts code by default — template provenance verification and malicious-code protection are weak. Do not install project templates from untrusted sources.
After opening a project, you will see a large run icon — click it and the workflow begins executing. You will be notified when execution finishes, and logs are available for inspection.
Long-chain workflows take a long time to run — this is normal. Unigen's foremost design goal is that quality never collapses; slowness is a consequence of today's limited compute.
With any project open, click the right edge to expand the Reflection Assistant (a project must be open), then enter:
/prism Give me a few short, hilarious ad concepts for facial tissues.
/prism does not persist anything, so you can directly compare Prism against the model's direct answer. For how it works, see Prism Reasoning (Section 10 is a verbatim transcript of exactly this example).
/plan From an idea to a complete novel
/plan persists its output and takes a very long time. For the current planner, see Loom: Reverse Deliverable Planning on a Graph Blackboard. You can then execute, improve, and publish the resulting workflow.
- Creators → Running Your First Workflow
- Developers → Workflow Development Overview