MiniMax H3 Prompt Tags (T8)

MiniMax H3 is picky about how you reference its inputs. The prompt doesn't say "use the reference video" - it says <Video 1>, <Audio 2>, <Picture 1>, and if those numbers don't line up with the inputs you actually wired in, the model can silently ignore a reference or pick the wrong one. The Prompt Tags node exists to fix that mess before it reaches the sampler: give it your prompt and your counts, and it normalizes the tags, validates the numbering, and hands back a prompt that matches your graph.

It's one of the few stable nodes in this pack - no "Advanced", no "Experimental", no fail-closed drama. That tracks with its job: it's the boring plumbing that makes the fancy stuff work.

How it works

The node takes your raw prompt and the number of each reference type you're actually connecting: picture_count, video_count, audio_count. It then normalizes tag casing and whitespace (<picture 1> becomes <Picture 1>), and checks that the tags you used are consistent with the counts and the ordinal settings.

The two inputs most people trip over are the audio ordinals:

• source_audio_ordinal - which <Audio N> refers to your source/input audio.
• prompt_primary_audio_ordinal - which <Audio N> is the prompt's primary audio reference.

Defaults are 1 for both, which is right for the common "one audio in, prompt points at it" case. They matter when you have more than one audio reference and the prompt's attention should go to a specific one.

Then there's strict (default true). In strict mode, ambiguity, missing tags, or duplicate labels are an error - the node will not silently guess. Flip it off and it gets lenient, which the README only really blesses for legacy single-reference workflows.

Outputs: a cleaned prompt and a report string (or OK) you can preview to see the warnings.

Why you'll hit "prompt media tag validation failed"

The pack's other nodes validate tags too, and the README's troubleshooting path is worth memorizing: when you see that error, check the task type first (T2VA vs I2VA vs Ref2VA expect different tags), then check that each tag's number actually matches a connected input, then check your counts here. Nine times out of ten someone typed <Video 2> with one video wired, or left audio_count at 0 while the prompt references <Audio 1>.

Installing

In the T8mars MiniMax H3 Audio T8 pack:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8

or ComfyUI Manager → search "MiniMax H3 Audio T8" → install → restart. No pip deps, no model downloads. The H3 weights, CLIP, VAEs, and LoRAs are yours to supply.

The take

This is the node you reach for when your workflow is getting multi-reference: a reference video plus a reference audio plus a first frame is exactly where tag numbering turns into a debugging session. Wire it in front of the conditioning node, set the counts to match your graph, and let strict mode catch your mistakes at prompt-build time instead of at render time. It won't make H3 obey you better - nothing will - but it removes one entire class of "the model ignored my reference" that's actually "I mislabeled my reference."