MiniMax H3 Enhance-A-Video Ref2VA / Hybrid Composer (Advanced EXP)

Ref2VA is the H3 workflow where a reference clip guides your generation - and the instinct to "enhance the whole thing" is exactly wrong there. If FETA scaled reference rows too, you'd be sharpening the thing you're supposed to copy, and your reference-guided output would drift. MiniMaxH3EnhanceAVideoReferenceComposerT8Advanced is the isolated composer for native Stock20 Ref2VA (and task-type Hybrid) conditioning: it computes FETA only from target-video Q/K, audits the exact reference-segment layout, and never directly scales reference, text, condition, or target-audio rows. Reference stays the source of truth; the enhancement touches only the generated video.

The model contract

This is the strictest gate of all the EAV composers. The model tooltip says it plainly: connect an unpatched native Stock20 H3 model - and LoRA, Prompt Relay, Sage, BlockCache, STG, Long Video, and model-Hybrid artifacts are all rejected. The composer needs a clean model because it must be the only thing touching the attention path; any other patch makes the audit meaningless. sigmas must be the exact 20-step Stock20 sigmas you send to the sampler.

The inputs you'll tune

mode - disabled is an exact bypass (the reference workflow runs exactly as stock), report_only audits FETA without modifying output, apply_exp turns the gain on. The tooltip has a nice trap warning: tau=0 is not an off switch. tau (default 4) is the experimental FETA weight - "a candidate, not an H3 optimum" - and setting it to 0 doesn't disable the mechanism. start_video_progress/end_video_progress bound where in the schedule FETA acts, max_workspace_mib (32) is the FETA score-buffer budget (not your workflow's total VRAM), and g_hard_limit (1.5) caps the observed gain.

Outputs

model (the composed model for the sampler), runtime (feed the EAV Audit node), and report_json. Standard EAV-family shape.

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 honest limits

The README is blunt here: a mechanical disabled/apply comparison on one 0.7MP material passed strict decode, but the Ref2VA enhancement path ran with a minimum VRAM margin around 417 MiB - below the project's own 512 MiB gate. So: don't stack this with other attention patches (it won't accept them anyway), don't assume 16GB is safe, and don't read the mechanical pass as a quality claim. Reference-only FETA is a clean idea; whether it's a better idea is still being tested - the composer just makes sure that when you test it, you're testing the mechanism and nothing else.