MiniMax H3 Hybrid Compatibility Audit / 混合模型组合兼容审计 (Advanced)

The Hybrid workflow is the fiddliest thing in this pack: you load a pruned quality base, stack a pruned reference overlay, add LoRAs, block cache, maybe SageAttention, and cross your fingers. When the model goes weird - warped reference layout, audio that doesn't match the picture, or just a silent OOM - you have no idea which layer did it. That's the job this node exists for. It isn't a generator, it's a checkpoint scanner: you drop it in the middle of the chain and it tells you, in one report, which of a dozen common failure classes your Hybrid setup has tripped.

It sits in the T8/MiniMax H3/Models/Experimental category, and the README's framing is worth taking seriously: this is an audit node from a pack whose author treats "will it still be true when a stranger runs it" as a design requirement. Everything about the Hybrid line is experimental, and the audit is the safety rail you're supposed to keep on while you experiment.

How it works

The mechanism is deceptively simple: it passes the exact same MODEL through untouched while auditing, in one pass, Hybrid offset-set integrity, LoRA ordering and overlap, Block Cache, SageAttention, Long Video and MultiKeyframe MODEL pairing, sampling, DynamicVRAM/VBAR policy provenance, current VRAM and host commit. You place it after every MODEL patch and sampler node and before BasicGuider - meaning it sees the final assembled state, not the intermediate one. Think of it as the last word before a render.

The enforcement input is the whole game. report_only (the default) never blocks and never mutates the MODEL - it just hands you the report. block_hard_conflicts actually raises on mechanically proven failures: wrong patch order, broken contracts, or a memory gate that can't be satisfied. Start in report_only. Only flip to hard-block once you trust the report, or you'll be fighting the node instead of the workflow.

The inputs that matter

• enforcement - report_only vs block_hard_conflicts. Default report_only; you rarely want to change it.
• minimum_current_headroom_mib (default 512) - the VRAM floor it expects to see before calling you safe. If you're on a tight card, this is where you feel the 16GB warning from the README.
• minimum_commit_headroom_gib (default 16) - same idea for system memory commit. On machines with 16GB RAM this default will flag you; that's the pack being honest, not broken.
• require_applied_vram_policy - only meaningful if you use the pack's VRAM policy nodes; it insists a policy was actually applied to the load path.
• positive (optional CONDITIONING) - connecting it lets the audit verify Long Video/MultiKeyframe MODEL pairing and reference-modality coverage.

Outputs

You get model (same MODEL, bit-for-bit intended), compatible (a BOOLEAN you can wire into a switch or just read in the UI), and report_json - the string you'll actually live in while debugging. Read it before the render, not after.

Installing it

This is the whole pack, so the install is the pack's install. ComfyUI Manager: search "MiniMax H3 Audio T8" and install, or run comfy node install minimax-h3-audio-t8. Manually:

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

then restart ComfyUI. The pack has no mandatory extra pip dependencies and doesn't auto-download H3 weights - you supply the diffusion model, CLIP, VAEs and LoRAs yourself. For Hybrid specifically you need the pruned FL2VA base and Ref2VA overlay that the Pair Inspector validates.

Common issues

Everything passes but the render is still broken. The audit only checks what it can mechanically prove - it won't catch a bad prompt or a wrong ref image. Trust compatible=false when you see it, but treat true as "nothing provably wrong," not "your shot is good."

block_hard_conflicts is blocking things you think are fine. Drop back to report_only, read the report_json, and fix the actual ordering violation. The hard block is honest: the pack's own regression notes repeatedly say a "compatible" check proves a chain executed, not that it looks good.