H3 Semantic Bridge / 应用条件 (T8 EXP)

This is the smallest node in the pair and the one people forget. MiniMaxH3SemanticBridgeApplyT8 takes a CONDITIONING and a bridge config, and hands back a CONDITIONING. Two inputs, two outputs, no widgets, nothing to tune. If you're looking for the settings - alpha, which model, token scope - you want the Config node instead. This one just does the work.

The split exists for a reason. H3 conditions are big objects with metadata attached, and the bridge rewrites their embedding rows through a 5120→512→512→5120 SiLU adapter. Keeping "which adapter, how strong" with the model settings and "apply it here" separate means you can drop the bridge into a working graph without unpicking anything.

Where it goes

Native H3 conditioning → Apply → your existing sampler or guider. That's it. Nothing else changes: same seed, same model, same prompt, same sampler - which is exactly what makes it a usable A/B test. Compare no-bridge vs original-model vs BUNNY, one variable at a time, and don't swap your text encoder in the same run or you'll be attributing two changes to one.

The conditioning output goes wherever the original conditioning would have gone. The report_json output is a string for diagnosis - which model file was used (by content hash), what the config looked like, how many times the bridge was applied. It's not scored, and it isn't a quality signal.

Note what the node does not do. It doesn't patch attention, doesn't add a LoRA, doesn't touch the audio branch, doesn't resize anything. The pack's docs are blunt that this is a condition-side adapter: existing models, sampling schedules, audio masks and upscalers are untouched.

The boring-but-important part: bypass

If the config you feed it has enabled off or alpha at zero, the Apply node returns your conditioning unchanged, without reading a model file and without adding receipt metadata to the condition. That's the intended off switch, and it means an empty or disabled Config isn't an error - it's a pass-through. It's also why the pack's advice is to leave configurations unset on stages where you don't want the bridge, rather than deleting nodes from a working graph.

The flip side: every active application records content identity and input/output tensor receipts into the conditioning's metadata, so a bridged condition carries provenance that an unbridged one doesn't. Change the model, alpha or scope and the cache contract for that bridge is invalid - the pack tells you to start a new chain id rather than reuse a cached result produced under different settings. That's not paranoia, it's the whole point of content-keyed caching.

Install

Same pack, no separate step. Manager → search MiniMax H3 Audio T8, or:

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

Restart ComfyUI completely; a page refresh won't pick up new Python classes. You also need the bridge weights in ComfyUI/models/semantic_bridge/ (see the Config node page) - the Apply node itself ships with the pack, but without a model it has nothing to apply. And the pack expects a current ComfyUI with native H3 support; if every node in the pack is red, update ComfyUI, the frontend and Manager, then restart from scratch before debugging anything else.

Where it goes wrong

Applied twice. The classic. The symptom is a condition that's been through the adapter twice, and the fix is removal, not addition: pull the extra Apply out and re-encode from the native conditioning node. If you're using Prompt Relay, don't place an Apply after the Relay at all - Relay carries its own internal bridge socket, and the node description says plainly not to double up. Same for the long-video in-node loops, which have their own per-pass bridge slots: those slots inherit the common config when left empty, and explicitly feeding one a disabled config bypasses the bridge on that pass without skipping the sampling for that pass.

Model missing. You'll get a file-not-found style error naming models/semantic_bridge. Refresh the combo list after adding files, and check the actual directory - the loader lists subdirectories, so t8_compat/ nesting is fine, but a file dropped into a random folder outside the enumerated roots won't show up.

"It made my clip worse." That's a real possible outcome and the defaults are set defensively for it: strength 0.10, and the author flags reference-audio and singing material as degradation-prone. The pack's own test protocol is worth copying - one seed, bridge off / original / BUNNY, listen to the generated audio all the way through the second segment, and don't peak-normalise the output to hide a quiet, noisy artefact. Original reference audio playing back clean tells you nothing about whether the model's generated speech survived.

Last thing: this pack has essentially no English discussion footprint, so there's no thread of people who already hit your exact problem. The repo docs are the troubleshooting resource.