MiniMax H3 Enhance-A-Video / FETA (Advanced EXP)

If you've got a MiniMax H3 clip where motion goes mushy or frames don't hold together, this is the node that tries to fix it inside the model instead of after the fact. It's an experimental port of the Enhance-A-Video technique - FETA, Feature Enhancement in Temporal Attention - into H3's actual attention pass. The idea, borrowed from the Enhance-A-Video paper, is to boost cross-frame interaction so the model keeps things temporally consistent while it generates.

Here's the mechanism, stripped down: the node computes the cross-frame interaction (CFI) from the target video's Q/K only - not the reference video, not the audio stream - and scales only the target-video attention output rows by a weight called tau (default 4). It does that in exact, low-memory chunks governed by max_workspace_mib, so the temporal score buffer doesn't eat your VRAM. It never touches audio attention, and it's not a post-process filter. It changes how the transformer attends across frames while sampling.

The inputs that actually matter:

• model - clean native H3 for the default stock20 profile. No LoRA, no cache, no attention patch: stock20 rejects every LoRA. For turbo8_alpha8, apply the corrected 208-module Alpha8 bypass LoRA before this node.
• sigmas - the exact SIGMAS your sampler receives. This gets checked strictly; a mismatch fails the run.
• mode - disabled is an exact bypass, report_only measures the CFI/gain without changing your output, and apply_exp actually applies the gain. Read that carefully: tau = 0 is not off. Off is disabled.
• tau - the enhancement weight. The tooltip is honest: 4 is "an upstream candidate, not a validated H3 optimum." Treat it as a starting point, not a setting you've earned.
• g_hard_limit (default 1.5) - a fail-closed guard. If the observed gain exceeds it, the node refuses the run rather than silently clamping and calling it paper-equivalent.

Outputs: model (into your sampler), runtime (H3_T8_EAV_RUNTIME - the pack's companion audit node consumes it so you can see exactly what happened), and report_json for inspection.

The part to internalize: this is an experiment, and the author says so on every panel. Automatic metrics only prove that trajectories and the joint audio changed - not that either got better. So the correct workflow is: run disabled and report_only (or apply_exp) with the same seed, and actually eyeball the results. That's not paranoia, that's the pack's whole design philosophy.

Where people get burned: stacking it with other attention patches. The README is explicit - don't pile standalone EAV on top of KJ Sage, BlockCache, STG, or Prompt Relay. If you want those combos, the pack ships dedicated composer nodes for them. And if you hit a "runtime contract" error, you fed it the wrong sigmas or a model with a LoRA in stock20 mode.

Install it via ComfyUI Manager (search "MiniMax H3 Audio T8") or:

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

Restart ComfyUI after. The pack has no forced pip dependencies - its requirements.txt literally reads "No extra packages" - and it never downloads H3 weights for you. You bring the H3 DiT, the Qwen/MiniMax CLIP, VAE, and any LoRAs yourself, and remember the H3 Community License itself geofences the weights out of the US, EU, UK and Korea regardless of what this GPL pack does.