TrainyThing Send

What it is

TrainyThing Send is the first half of TrainyThing Review, alone: the batch becomes a review queue on your phone, the node shows the link and a QR, and then the run finishes. No waiting, no held queue, no opinion about what ComfyUI does next.

That sounds like a downgrade until you've had a Review node sit on the canvas for forty minutes with ComfyUI frozen behind it. Send is the way out: queue the graph as many times as you like with the same slug and the phone queue grows while you keep generating. Review it whenever, then come back with TrainyThing Load or wire Wait if the split belongs in this run.

Both halves are the same machinery: Send owns the upload and the QR, Wait owns the polling and the slicing. When reviewing can happen on its own schedule, Send alone is the right tool.

How it works

Send builds the queue's configuration from its widgets, creates a queue (or appends to one that's still open), turns each image into a card, and uploads the PNGs to a local TrainyThing daemon. The uploads keep their prompt and workflow chunks - the same ones Save Image writes - so a keeper dragged back out of output/trainything/<slug>/blobs/ reloads the entire workflow that made it.

Before uploading it also walks your graph backwards from images: VAE Decode, sampler, text encoders, LoRA loaders, checkpoint. That fills in each card's seed, prompt and sampler settings so the verdicts come back actionable. It's best effort - custom samplers and loop packs stop it - and meta overrides it.

Then it hands back a handle and gets out of the way. The QR is drawn as an image preview with the link printed underneath, because ui.PreviewText doesn't render on custom nodes in 0.39.0.

Connection details are never widgets: TRAINYTHING_URL/TRAINYTHING_TOKEN from the environment, then ~/.trainything/config.json, then http://127.0.0.1:4747. A token widget would end up baked into every PNG and every shared workflow file, which is exactly how you'd leak it.

The inputs and outputs that matter

Send takes Review's inputs minus latent, timeout_minutes and on_timeout - there's no wait, so nothing to time out. The ones a beginner sets:

• images - the batch, one card each.
• slug - the queue's name in the link. Empty derives one from title; random gives an unguessable one. Use an open queue's name and your new cards are appended - that's the whole "queue the same graph ten times" trick.
• mode - single, pair (wire images_b, same batch size), choose or select (both need options, written id:Label, id:Label).
• public - opens the tunnel before any upload so the link works off your Wi-Fi. A failed tunnel gives you a console line and the local link, not an error.
• min_reviews, feedback, identity, order - opinions per card, when a reason is demanded, whether reviewers leave a name, and whether cards run in batch order or fewest-reviews-first.

Three outputs: queue (a TRAINYTHING_QUEUE handle - URL, slug, queue id, link, this run's card refs; only Wait takes it), link and slug (STRING, for a preview or for pasting into Load later).

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/ryanrinkel/comfyui-trainything
cd comfyui-trainything
python -m pip install -r requirements.txt   # ComfyUI's own python, or python_embeded\python.exe
# restart ComfyUI, then once:
npx -y trainything dashboard

Manager: search the pack title "TrainyThing", or comfy node install trainything. Deps are segno (QR codes), pillow, numpy; no model files. The daemon is separate and needs Node ≥ 22.5 - that npx line starts it in the background. The pack is built on ComfyUI's V3 node API and pins requires-comfyui = ">=0.39.0".

Where people get burned

The cache eats your second send. ComfyUI caches an unchanged Send, on the logic that those images are already on the phone. Queue the identical graph twice and nothing new arrives - change a seed or any widget to force a real send.

The queue output isn't a value you can feed anywhere else. It fits Wait's socket and nothing else. If you're not waiting in this graph, ignore it and use the slug.

Appending has rules. A queue made by the CLI, the dashboard or the MCP server keeps its settings: a different mode is an error naming both, different options give way to the queue's, a queue with no min_reviews adopts the node's.

No daemon, one line. no TrainyThing at http://127.0.0.1:4747; run npx -y trainything dashboard once. The node tries to start a local daemon once, with a 15-second budget; if that fails, start it yourself.

options is fussy in choose/select, on purpose. Lowercase letters, digits, - and _, 2 to 12 of them, skip reserved. A malformed list errors with the offending entry named.