What's New in TongFlow — Everything Since v0.1.2

Since open-sourcing TongFlow, we’ve shipped seven releases in two weeks. New models joined the catalog, new node types landed on the canvas, and the way you connect nodes got a rethink. If you’ve been away since v0.1.1, here’s everything that changed — grouped by what it means for your workflows, not by version number.

New to TongFlow? It’s an open-source multi-modal GenAI workflow studio: every AI model is a node on an infinite canvas. Start with the introduction for the full story.

New models and plugins

The official plugin catalog kept growing across these releases:

• TripoSplat — single image → 3D Gaussian Splatting, running on GPU via Modal. We wrote a dedicated post on this one. (v0.1.2)
• SCAIL-2 — end-to-end controlled character animation: drive a single character image with any reference video. Also has its own post. (v0.1.2)
• Unlimited-OCR — Baidu’s long-horizon document / PDF → text on the parse-document slot, a GPU alternative to Docling and PaddleOCR. (v0.1.3)
• Doubao Seedance 2.0 — a Volcengine Ark API plugin covering text → video, image → video, first/last-frame video, and image + audio → video. (v0.1.8)
• Bernini-R — a unified renderer plugin backing the new video-editing node. (v0.1.8)

New things you can build

• 3D on the canvas — the 3D model node now renders Gaussian-splat assets (.splat, .ply, .spz, .ksplat, .sog) in-app via SparkJS, with a free-tumble trackball camera and on-demand rendering that stays smooth even on large splats. Plugin model outputs are persisted and shown on the canvas instead of being dropped. (v0.1.2)
• Images → Video — a new multi-image reference node. Connect 2–9 reference images plus a prompt, and Seedance fuses them into a new video. Reachable from the smart-island compose menu when the right nodes are selected. (v0.1.8)
• Video editing — a new video-edit node for instruction-driven edits, backed by the Bernini-R renderer. (v0.1.8)
• Resolution tiers — image generation nodes gained a 1K / 2K / 4K picker alongside the aspect-ratio control. The tier scales the aspect ratio’s base dimensions, with no new contract field to learn. (v0.1.6)

A cleaner canvas

Connecting nodes got a rethink in v0.1.7:

• Reconnect-only edges — you no longer drag from a handle to create a wire. Connections are created through the operation panel (expand / compose), which keeps the graph intentional and tidy. You can still reconnect an existing edge’s endpoint to another node.
• ABI-validated reconnects — dragging an endpoint is checked against the contract: the upstream modality must match the target handle, single-value handles accept only one edge, add nodes fan out to a single edge, and modality nodes can’t feed each other. Drop an endpoint on empty canvas and TongFlow offers to delete the connection.
• Aspect-ratio picker labels now wrap instead of clipping past the button border in English.

Robustness fixes

• Windows non-UTF-8 locales — plugins no longer crash on systems whose ANSI code page isn’t UTF-8 (notably Simplified-Chinese Windows on GBK / cp936). A spawned Python printing a non-ASCII character like ✓ used to raise UnicodeEncodeError and surface as “downloading weights failed”. Every Python spawn now forces UTF-8 mode. (v0.1.4)
• Desktop release pipeline — per-arch installers are now uploaded to a draft release that’s published once, fixing failures under GitHub’s immutable releases. (v0.1.6)
• Image Fusion — a prompt fed from an upstream text node is now actually used at execution, instead of the node refusing to run with “required input text is empty”. (v0.1.8)

Get the updates

Grab the latest desktop build from the Releases page, or git pull if you’re running from source. New plugins install from the in-app plugin manager and are usable immediately — no restart.

As always, the official GPU/CPU plugins run on Modal’s free tier, so you can try every one of these without owning a GPU. A star on GitHub helps others find the project, and Discord is the place for bug reports, ideas, and showing off what you’ve built.

More models, more modalities, more ways to combine them — keep building.