YAW Settings Importer (drop a workflow here)

The problem this solves isn't the one you'd guess

If you run the YAW template family - the author's Wan, Bernini, MiniMax H3, LTX and Krea workflows that get reissued every few weeks - you know the annoying part. v0.39 becomes v0.50, v0.50b becomes v0.51, and the new template is a different graph. You liked your prompts, your LightX2V stack at 0.7, your 6 steps at CFG 1. The new version starts fresh.

The usual fix is opening the old workflow in a second tab and retyping everything, which is how people lose a LoRA's on/off state or a second strength they'd forgotten about. YAW Settings Importer is one node you drop that old workflow file onto: it reads both graphs, works out which settings correspond, and writes them into the workflow on your canvas.

How it actually works

No hidden link trickery. The frontend extension (web/settings_importer.js) reads your dropped file, grabs the serialized live graph too, and POSTs both to /sampling_planner/import_settings. The server side (settings_importer.py) is pure Python with no ComfyUI imports - it locates semantic slots by node type, plus the titles the YAW templates have kept across versions, plus wiring hints - and returns a patch list the browser applies.

That's why cross-version import works when a naive copy wouldn't: the server doesn't trust widget order, so it finds the sampler that's actually live rather than a muted branch or a widget behind a linked selector. And it refuses to guess - anything missing or ambiguous is skipped with a reason, which is what the report is for. The whole application is bracketed in the graph's change hooks, so Ctrl+Z undoes it.

The six checkboxes, and which ones you'll touch

Everything here is a boolean widget. No input sockets and no outputs at all - RETURN_TYPES is empty, it isn't an output node, and nothing is wired to anything. It never runs generation.

• prompts (on) - positive and negative prompts.
• loras (on) - compatible Power LoRA rows: High/Low, main, or upscale. Uses rgthree's row factory; if the target's loader version doesn't expose it, your existing rows are kept and reported.
• sampler_settings (on) - sampler/scheduler, step budgets, CFG, shift, acceleration/priority, clip length.
• task (on) - matches T2V/I2V by muting and unmuting the groups titled T2V/I2V, the same thing the templates' Fast Groups Muter does. H3's I2V-only and End Image groups are handled too.
• seed (off) - correctly off by default: you want the settings, not the exact clip.
• accelerator_lora (off) - copies LightX2V / H3 Turbo / distilled names and strengths, but only when a unique compatible loader exists. Linked strength controls come back as "review manually," which is the right answer rather than an error.

That last one earns its checkbox. Distilled acceleration is exactly where people get burned - the category wants CFG 0–2 and the step count it was trained for, and mixing accelerated and base values silently gives you oversaturated mush.

Click the dashed square any time to reopen the last report: "✓ 6 copied · 2 skipped", plus reasons.

Install

Manager: search YAW Settings Importer and install. Manual, from custom_nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/boobkake22/ComfyUI-YAWSettingsImporter.git

Restart and refresh the browser. The pack declares no dependencies, no models, no API keys. It does need whatever node packs make up your YAW templates - the importer only moves values between nodes that already exist on your canvas.

Where people get burned

• You dropped an API-format prompt JSON. No node titles, nothing to anchor to. Export the workflow itself, or drop a ComfyUI-saved PNG.
• PNG/WebP only if ComfyUI wrote it. The graph has to be embedded in the file; a screenshot or re-encoded image is a dead end - the same metadata your image host strips on recompress.
• For a video, drop the paired image. The README's own advice, and it follows from the above.
• You're inside a subgraph. Import targets must be the root graph, and the node tells you so rather than quietly editing the wrong one.
• It won't flip your presets. Modern-family imports write the active target controls; they don't switch Turbo, Basic/Advanced, upscale or DMD groups. Differing presets get reported, not enforced. A manual sigma schedule likewise only copies when the target runs one too.
• Family crossing. LoRA rows don't transfer between recognized model families, and T2V/I2V→R2V keeps the target's generation mode while reporting the task mismatch.
• Don't touch the graph mid-import. It snapshots your workflow and refuses if it changed in flight, which beats half-applying a patch list.

Verdict

Narrow, and good at its narrow thing. Corpus chatter about this pack is close to zero, so expect the on-node report to explain itself rather than a forum thread. If you live in the YAW templates, this turns a version upgrade into a drag and drop. If you don't, there's nothing here for you and it won't pretend otherwise.