MiniMax H3 Enhance-A-Video + Prompt Relay (Advanced EXP)

Here's a failure mode the README calls out directly: two of T8mars' standalone nodes both want to own the same attention entry point, and plugging them both in makes them fight. Prompt Relay patches attention to route local events; Enhance-A-Video (FETA) patches it to compute cross-frame interaction. Stack them naively and one silently wins, or neither works as intended. MiniMaxH3EnhanceAVideoPromptRelayComposerT8Advanced is the combined owner that does both in one authenticated wrapper: Relay performs its local-event attention route first, then FETA scales only the target-video output rows. And crucially - it does not add model forwards. You get both mechanisms for the same number of forwards the Relay alone would have cost.

How to wire it

model connects to the MODEL output of Prompt Relay Conditioning - the tooltip says apply_exp wants at least two active events. If you're on the turbo8_alpha8 sampling profile, apply the corrected Alpha8 bypass LoRA after Relay and before this composer - the pack's wiring rule from the README is explicit about that order (LoRA before the Relay gets actively rejected). sigmas is the exact SIGMAS sent to the sampler. Then the standard EAV knobs: mode, tau (4), progress window, max_workspace_mib (32), g_hard_limit (1.5), and sampling_profile (stock20 - the bound visual/reference task - or turbo8_alpha8, which is strictly limited to T2VA with the corrected 208-hook bypass LoRA).

Modes

disabled preserves Prompt Relay exactly and disables only EAV (that's the "undo the fight" switch). report_only measures FETA after Relay without scaling - the safe first run. apply_exp turns the target-video gain on. Outputs: model, runtime (EAV Audit), report_json.

The honesty caveats

The README's notes on this composer are worth reading before you queue anything: one 736×416×124, 20-step mechanical baseline passed strict media checks, but the VRAM margin landed below the project's 512 MiB gate. So this is not a stable-quality claim, not a 16GB-safety claim, and not an "audio is non-worse" claim. It is the only sanctioned way to run Relay + EAV together - because doing it without the composer is exactly how you get silent attention-patch interference.

Installing it

From MiniMax H3 Audio T8 (T8mars/T8star, GPL-3.0). Manager search MiniMax H3 Audio T8, or:

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

restart. No forced pip deps; models to models/ by hand. H3 weights' Community License excludes the US, EU, UK, South Korea.

The trap

If you connect a plain model (no Relay) and expect Relay behavior, you'll get silence where local events should fire - this composer is the Relay, it doesn't read one from elsewhere. And sampling_profile isn't decorative: pushing turbo8_alpha8 beyond T2VA is explicitly unsupported, and the corrected Alpha8 bypass LoRA is mandatory for that profile. Wire the profile, wire the LoRA in the right order, and don't stack a second attention patch on top - the composer is meant to be the only one.