MiniMax H3 Official Risk Diagnostic / 官方问题证据诊断 (Advanced EXP/T8)

Somewhere between "my H3 render looks wrong" and "my H3 render is objectively broken" there's a huge grey zone, and this node exists to stop you from guessing through it. The Official Risk Diagnostic is a read-only evidence classifier for the failure classes T8's pack keeps tripping over: VRAM pressure and V-copy stalls, reference loading, Sage FP8 attention, audio numerical integrity, multi-speaker cross-talk, and that delightful periodic-dark-flash artifact. You feed it facts, it tells you which of those it can actually implicate - and if it can't, it says so. "Unknown remains unknown" is the honest core of the design, not a cop-out.

How it works

It's a pure classifier with zero side effects. You describe the render you're about to run (or just ran) and optionally paste in the JSON reports that other T8 nodes already produced. The node combines those signals against its risk catalogue and returns a status string, a risk count, and a full JSON report. No model fingerprint is checked, no canvas-size hard gate is applied - the author deliberately refuses to block on "this resolution is too big," because 1920x1088 is a reference area, not a law.

The inputs that matter:

• width, height, length - your canvas and frame count (32-aligned width/height; H3 frame counts typically land on the 17n+5 rhythm, like 22 or 124, so use the real number from your render).
• reference_media_count, speaker_count, isolated_voice_reference_count - the reference load and the multi-speaker cross-talk risk live here.
• attention_backend - stock, sage, sage_fp8, sol, other, or unknown. This is where the Sage FP8 risk gets classified, and it's the one Combo most beginners get wrong because they don't actually know what their install is running.
• runtime_report_json, audio_report_json, frame_report_json - optional, paste the outputs from other T8 diagnostic nodes and it gets a lot more specific.

Outputs are status (STRING), risk_count (INT), and report_json (STRING). There's no image or video output at all - this is a "sit next to your workflow and read it" node, not something you wire into your video path.

Why you'd actually use it

T8's own troubleshooting docs are blunt: "根据启动终端的第一条报错判断" - judge by the first error in the startup terminal. This node is the in-workflow version of that habit. If a render comes out with dark flashes or garbled audio, run the diagnostic with the render's real parameters instead of blindly dropping resolution or swapping samplers. It's also genuinely useful as a preflight: the 21-community-advanced "System Cache and Diagnostics" workflow wires it in read-only, so you can see risk posture before committing a long render.

Installing it

It ships inside the comfyui-minimax-h3-audio-T8 pack, so:

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

or search "MiniMax H3 Audio T8" in ComfyUI Manager. Either way, update ComfyUI itself first - this pack leans on native MiniMax H3 support (comfy.ldm.minimax, comfy_api.latest, comfy.weight_adapter), and a stale core makes every T8 node go red at once. The base requirements.txt is intentionally empty; dependencies come from ComfyUI.

Where people get burned

The trap is expecting it to fix things. It won't unload models, change preview settings, or patch your attention backend - it only classifies. And because it's an experimental node with zero search impressions behind it, there's no folklore yet; the README's rule applies: unknown evidence stays unknown, so a clean report means "nothing this node knows about is implicated," not "your render is definitely fine." When it reports a risk you don't recognize, treat it as a lead to investigate, not a verdict to argue with.