Introduction

TongFlow is a multi-modal AIGC studio that runs on your own machine. You build creative workflows on an infinite canvas — drop materials onto it, transform them between modalities (text, image, video, audio, 3D), and combine the results.

The whole project is open source on GitHub at tong-io/tongflow under AGPL-3.0. This is v0.1.0 — early days, small surface, no commercial sleight-of-hand.

Core ideas

  • All models are modality transforms. A text-to-image model is text → image. A speech recognizer is audio → text. A 3D generator is image → 3D. TongFlow wraps every model as a node with typed inputs and typed outputs on the canvas.
  • All modalities are first class. Text, image, audio, video, document, URL, and 3D model all live on the same canvas.
  • The interface is three verbs. Add materials. Transform between modalities. Combine results. No complex parameter panels.

Local-first by default

  • Workflows and uploaded materials live in a local SQLite file (data/tongflow.db) and on local disk (data/uploads/).
  • No TongFlow account, no central CDN, no telemetry.
  • AI inference uses two external services that you configure yourself:
    • Modal for GPU/CPU workers (their free tier — $30/month — includes meaningful H100 time).
    • One LLM provider of your choice: OpenRouter, Gemini, OpenAI, or DeepSeek.

You bring the API keys. We never see them.

What’s in the box

  • Seven add types on the canvas (text, image, audio, video, document, URL, 3D model)
  • Transform nodes across all five modalities
  • Combine nodes (image fusion, lip sync, voice cloning, character swap, motion transfer)
  • Named backend models: Z-Image, FLUX.2 Klein 9B, LTX-2, SeedVR2, Gemma 4, Qwen3, ACE-Step
  • One Docker command to self-host

For the precise list of what works (and what is still pending), see the README — it is the source of truth.

What this is not

  • Not a hosted SaaS with SLAs and concurrency tiers (yet). app.tongflow.com exists as a preview convenience.
  • Not a no-code black box. You arrange nodes; you understand what each one does.
  • Not “any model, any time.” We name the models we ship and link the providers we depend on.

Next steps

  • Getting started — install in one Docker command, configure env, run your first workflow.
  • Interface overview — the canvas, Smart Island, left sidebar, mode switch.
  • Node types — the actual catalog of add / transform / combine / helper nodes.

If you want to extend TongFlow with your own models or new node types, see docs/feature-registry.md and docs/plugins.md in the repo.