🏝️ Dependency Generator

Every ComfyUI regular has hit the same wall: someone shares a workflow, you load it, and half the nodes are red because the checkpoint, the LoRAs, and the VAE aren't on your machine. The workflow travels fine - the assets don't. Shima's Dependency Generator exists to fix the sharing half of that problem: it scans the graph on your canvas and produces a JSON manifest of every model, LoRA, VAE, and ControlNet the workflow depends on, ready to hand to its sibling, the Dependency Installer.

Mechanically it's a prompt audit, not a downloader. The node receives the current execution graph (via the hidden PROMPT input ComfyUI provides), walks every node, and matches known loader types: CheckpointLoaderSimple, CheckpointLoader, GGUFLoader, LoraLoader, VAELoader, ControlNetLoader, plus Shima's own ModelCitizen/PanelModelCitizen/PanelBNDLer multi-loaders. Each hit becomes an entry in the output JSON with the asset name, a type label (Checkpoint/LoRA/VAE/ControlNet), and a placeholder download URL you're meant to fill in before sharing. It also dedupes entries and skips obvious placeholders like "Baked VAE".

The one input, scan_now (BOOL, default on), is your trigger - a scan fires when the graph executes. The single output, json_stub (STRING), is the manifest as a JSON string. Wire it into the Dependency Installer's json_stub input, or copy it into a text file next to your workflow PNG.

A couple of honest caveats. First, this is a widget-based scan: resolve_value deliberately returns None for values that arrive over wires, so assets picked by an upstream selector node won't be caught - the docs describe it as a deep scan, but the version you're running reads widget values. Second, the URL field comes out as https://INSERT_URL_HERE, which is you-shaped, not automated: the generator tells you what you need, and you (or the Installer) still supply where to get it. And it's an output node, so it runs on every queue and prints the JSON to the console - harmless, just don't read the console noise as an error.

Where it earns its keep: publishing a template or island to a community, where a manifest text file makes the difference between "here, load this" and "here, load this - oh and you'll need four models, two LoRAs, and a VAE." It's the same sharing culture as embedding workflows in PNGs, just extended to the parts that can't be embedded.

Install the pack via ComfyUI Manager (search "Shima") or:

cd ComfyUI/custom_nodes
git clone https://github.com/KDB-USJP/shima_wf

then restart. No models to download - this node's whole job is describing which models a workflow needs.