MiniMax H3 Prompt Queue (TJ)

If you're generating a MiniMax H3 video, you usually want several clips - one per prompt, one per shot, one per line of a storyboard. Clicking Queue, waiting two minutes, changing the prompt, clicking Queue again gets old fast. MiniMax H3 Prompt Queue (TJ) is the node that automates that loop: up to 20 prompt slots, and it advances to the next one and re-queues itself every time a run finishes.

Here's the mechanism, because it's the interesting part. The Python side is deliberately dumb - it just exposes the slots and returns whichever prompt matches the current index. All the "run finished, submit the next clip" logic lives in the pack's frontend JS (web/prompt_queue_tj.js), which watches the executing/status websocket events and, when a run completes, calls app.queuePrompt() again with the next index. Crucially, it doesn't monkey-patch queuePrompt - it just calls it - so it won't collide with Multi Switch's queue hook if you use both in one workflow. The node also forces itself to always re-run (IS_CHANGED returns NaN), because the JS is mutating current_index between runs and ComfyUI would otherwise happily serve you a cached stale prompt.

The inputs you'll actually touch:

• prompt_count - how many of the 20 slots are active (1–20).
• current_index - which slot is live right now; the JS bumps this as it goes.
• current_queue - a position counter the JS uses to track the loop.
• prompt_1 … prompt_20 - the multiline prompt texts themselves. Most people never use more than a handful.

Outputs are what make it a control hub rather than just a text box:

• prompt - the active prompt string. Wire it into a sampler/sequencer.
• index - the 1-based clip number.
• total_count - the total number of prompts.

In the H3 workflow, index and total_count are the glue: TJ_H3_Output uses them to know when it's seen the last clip of an One-Take sequence (and should stitch), and TJ_H3_LoadLatentCheckpoint uses index to know clip 1 from clip 2 (the first clip tolerates a missing checkpoint file, later ones don't).

One honest caveat: "automatically keeps queuing" means the queue runs until it's done, whether or not you're watching. On a long H3 shoot that's a lot of GPU-hours if you forgot to fix a prompt. Know your loop, and use the queue count or the Stop button when you step away. It's also worth noting this node isn't H3-exclusive - nothing in it knows about video - but in practice it exists to drive the H3 nodes in this pack, and that's where it shines.

Install

Part of ComfyUI-TJ_NODE. ComfyUI Manager → search "ComfyUI-TJ_NODE", or:

cd ComfyUI/custom_nodes
git clone https://github.com/designloves2/ComfyUI-TJ_NODE

Then restart. The queue logic is frontend-only, so no extra Python packages are needed for this node.

Troubleshooting

• It stops after one run - the pack's web/ JS isn't loading. Check the browser console for a JS error, and confirm you restarted ComfyUI (frontend files are picked up at startup) rather than just reloading the page.
• You get the same prompt every clip - stale cache; update the pack, since the always-re-run fix is version-dependent.
• The whole pack feels heavy for this - it is a big pack. You're not wrong; but the H3 loop needs it, and there's no standalone version of this node that I'd recommend over it.