Toonflow review Canvas connections and asset retention
Pinned local checks of media connections, asset cleanup candidates and Chinese script decoding, with deployment requirements and generation limits.
AgentIndex · Published 2026-11-05 · Updated 2026-11-05
Toonflow organizes scripts, storyboards, character assets and generation tasks on an infinite canvas for AI short dramas, motion comics and short videos. It is relevant to creators who already plan to use image or video models and need to manage relationships between shots and materials.
Our local checks found useful structural safeguards: selected code rejected incompatible media ports and retained referenced asset directories when calculating cleanup candidates. That supports a limited workflow trial. We did not start the full client or generate a video, so this review does not establish finished-film quality.
Pinned source and twenty local checks
We fixed commit 6086cdffceb704a632eca143dd16cd461a78cdac on November 5, 2026. The repository declares Bun 1.4.3. We ran selected modules using Node.js 26.6.0 native TypeScript stripping, with a local Axios dependency for the file module's error recognition. This is not a test of the full Bun application. Declared runtime and workspace structure
| Check | Observed behavior |
|---|---|
| IMAGE to IMAGE and wildcard types | Accepted |
| IMAGE to VIDEO | Rejected |
| Self connection, absent port, receiver veto | Rejected |
| Referenced asset directories | Excluded from cleanup candidates |
| Unreferenced valid node directory | Returned as a candidate |
| Root and nested canvas markers | Root recognized; nested marker rejected |
| UTF-8 and BOM-marked UTF-16LE Chinese | Decoded |
| Binary text with a null byte | Rejected |
Twenty checks passed, including one assertion about the declared Compose loopback binding. The reproduction fixture records each check and the inspected source hashes. No UI interaction or complete production workflow was included.
Connections enforce declared media types
A canvas cannot safely treat an image, video and text value as interchangeable. Our IMAGE output connected to an IMAGE input, while IMAGE-to-VIDEO matching failed. An absent handle, self connection and explicit receiver veto were also rejected. Connection implementation
This can help surface a structural mistake before a generation request. It does not tell you whether the reference image depicts the right character, whether the provider understands the shot description or whether generated clips preserve visual continuity. Those require a real production sample with the selected model.
For a first trial, use one short shot with known input material. Inspect its declared ports, parameters and output file before expanding to a sequence. With custom nodes, review wildcard declarations and additional receiver rules. A permitted connection means the declared structural conditions passed, not that the creative result will be usable.
Asset references and recoverable projects
Shared materials make node deletion consequential. The tested function examines retained nodes and their nested assets/… references when calculating directory cleanup candidates. In our specimen, both a referenced shot directory and a referenced shared directory were retained. An unreferenced valid node directory was returned. Asset-directory calculation
The result is a list of proposed paths. We did not delete files or establish that every concurrent edit preserves assets. Before upgrading or moving a project, preserve configuration, project files and materials together. A canvas JSON file alone should not be assumed to contain everything needed for recovery.
File recognition also has a narrower purpose than validation. A leading root toonflowCanvas:true marker takes a fast path. We confirmed that a file with that prefix and malformed remaining JSON can still be recognized. A nested marker was rejected, and a marker appearing after a long first field was found through the full-read path. Recognition therefore should not be treated as complete structural validation of an imported project.
Test recovery with a disposable copy: open it, inspect referenced materials and exercise the intended export path. Our selected module checks do not establish export completeness or a successful backup restoration.
Chinese scripts and generation dependencies
UTF-8 and BOM-marked UTF-16LE Chinese samples decoded correctly; a binary sample containing a null byte was refused. DOCX filename recognition worked, but recognizing an extension is separate from extracting a Word document. Document and text handling
Keep an original script when testing imports. Check paragraph boundaries, punctuation, character names and the resulting shot breakdown before sending material to a model. Text decoding, document conversion, story development and video generation are distinct stages. Success at one stage cannot stand in for acceptance of the sequence.
The project documents third-party API configuration and integration with local ComfyUI and language models. Its MIT license permits self-hosting, while model access, compute and provider charges remain separate requirements. Free application code does not establish free image or video generation. Project capabilities and cost conditions
Deployment and adoption
The checked Compose file declares a 127.0.0.1:3000 binding and a persistent volume for /app/data. The README states that the business UI and API lack separate login authentication, so remote use needs access restrictions or an authenticated reverse proxy. We inspected these declarations; we did not build the Docker image or test remote deployment. Compose configuration
Choose a small project with identifiable characters, a few shots and traceable references. Confirm asset handling, recovery after a failed task and actual model output before considering batch production. Teams requiring demonstrated film quality need their own trial with the intended provider and material.
To retain production decisions across agent sessions, see the Obsidian Second Brain Crew review. For editable diagrams as workflow artifacts, continue with the Next AI Drawio review.