MiniMax H3 Enhance-A-Video + STG (Advanced EXP)

This composer pairs the Enhance-A-Video / FETA temporal boost with a second idea: skip-block guidance, STG for short. The pitch is that you get both techniques in one owned node instead of stacking two third-party patches and hoping they don't fight. Like the other EAV composers in this pack, it's Stock20-only, experimental, and built around auditing exactly what happened rather than promising a quality win.

How the mechanism works: STG runs a weak branch of the model alongside the main one - the same as guidance, but instead of a separate model you skip a set of H3 "double blocks" in the weak path and use the difference as extra guidance. stg_double_blocks (default "25") is the comma-separated list of blocks skipped only by the weak branch, stg_scale (default 0.35) is how hard that guidance pushes, and stg_start_progress / stg_end_progress (0.25 → 0.85) bound when it's active. Meanwhile FETA is applied consistently to both the main and weak branches, and the runtime audit verifies the exact number of extra joint audio-video forwards so the "extra compute" isn't a mystery.

The subtle knob is mode, and it's easy to misread:

• disabled - keeps STG active but disables only EAV. This is your controlled baseline: pure skip-block guidance, no FETA.
• report_only - measures FETA on both branches without applying the target-video gain.
• apply_exp - the full thing: STG plus the FETA gain on the target video.

That's a deliberate quirk: "disabled" here doesn't mean the node is off, it means EAV is off. If you want literally nothing, don't connect the node at all.

The fields that bite beginners:

• sigmas - exact native Stock20 SIGMAS from the sampler. Contract-checked.
• shift_video (default 12) - must match the native H3 video clock your schedule uses.
• rescale - must stay 0. It's not validated for this combo, and the tooltip is a warning dressed as a description.
• tau - the usual experimental FETA weight, 4 as a candidate not a conclusion.

Outputs are the standard composer set: model (into the sampler), runtime (H3_T8_EAV_RUNTIME), and report_json. The audit matters here because STG + FETA means extra joint AV forwards, and the node is transparent about counting them - the README's verdict is that it passed low-load deterministic contract checks and workflow import, but was not stress-tested, and it's not being marketed as "better quality, faster, or safer on 16GB." That's the whole tone of this pack, and it's worth respecting: run it, read the report, judge the video yourself.

Install: ComfyUI Manager, search "MiniMax H3 Audio T8", or clone it:

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

Restart after. No forced pip dependencies and no auto-downloaded weights - H3 model, CLIP, VAE and LoRAs are yours to supply. And before you reach for this: make sure your setup is clean. The tooltip is explicit that you shouldn't stack this on top of standalone STG, BlockCache, Prompt Relay, Sage, or Long Video hooks. One guidance owner per graph.