MiniMax H3 Hybrid Artifact Builder / 小型混合补丁构建 (Advanced)

The 09-hybrid-model corner of this pack researches "hybrid" H3 models - combining FL2VA and Ref2VA task behavior - and this node is the storage side of that research. Its job: take a hybrid plan and build (or reuse) a small FP16 "target-slice" artifact, 13.8–83.1 MiB, saved under ComfyUI/models/h3_hybrid_artifacts. The two rules that define it are in the description: it never copies a complete model, and it never overwrites a mismatched artifact.

Here's the idea. Rather than fusing or loading two full H3 models to test a hybrid idea, the pack stores only the slice of weights that a hybrid plan needs - a few tens of megabytes instead of ~42GB. The artifact is content-addressed, meaning its identity and filename are derived from the plan's content, so the same plan always maps to the same artifact path. Build once, reuse forever: if the artifact already exists and matches, the node returns it without rebuilding. If the artifact exists but doesn't match the plan (stale or tampered), it refuses to overwrite it - you go to the Maintenance node to sort that out.

The input surface is tiny: a single hybrid_plan (H3_T8_HYBRID_PLAN), which you get from the pack's hybrid plan/loader nodes. Outputs: hybrid_artifact (the H3_T8_HYBRID_ARTIFACT object the hybrid loader consumes), artifact_path (so you can see exactly where it landed on disk), and report_json (including the content-address, so you can verify reuse).

Who is this for? Honestly, a narrow group: people experimenting with FL2VA/Ref2VA hybrid models and wanting reproducible, disk-cheap patches. If you're not building hybrid research workflows, this node does nothing for you - it has exactly one required input and no optional ones, and it won't do anything interesting until you have a plan. But if you are doing that work, the value is real: content-addressed, collision-checked, minimal-disk-weight artifacts with no full-model copies, and a safe no-overwrite contract so a mismatched patch can't quietly corrupt your hybrid builds.

The caveat in the pack's own voice: this is experimental research infrastructure, and it's about storage hygiene, not quality. Building an artifact proves the slice was built consistently - it doesn't prove the hybrid idea is good, or that the result is 16GB-safe, or that it beats either base model. The README's hybrid-model section is explicit that these are "research" workflows with no quality claims attached.

Install: ComfyUI Manager → "MiniMax H3 Audio T8", or:

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

Restart after. No forced pip deps, no downloads - the artifacts live under ComfyUI/models/h3_hybrid_artifacts and the base H3 models are yours to supply.