MiniMax H3 Hybrid Model Loader / 混合模型加载 (Advanced)

The whole point of the Hybrid line in this pack is that you don't have to run one giant model to get both first/last-frame (FL2VA) and reference (Ref2VA) behavior. The trick is a pruned quality base plus a small reference overlay, and the order you apply things in is not cosmetic. This loader is the node that enforces that order so you don't end up with a patched-to-hell model you can't trust.

It does one clean job: it loads the quality base through ComfyUI's stock diffusion-model loader - the same path a normal checkpoint uses - then applies only the connected small hybrid artifact to a clone of that model. The base loader path matters because it preserves ComfyUI's DynamicVRAM / VBAR file-backed behavior. If you hand-load the base yourself and mangle the loading path, you lose the memory bookkeeping that keeps a 33B-class model on a consumer card at all.

How it works

Under the hood the node splits loading from patching. The stock loader reads quality_base and hands back a clean MODEL. Only then does the hybrid_artifact (an H3_T8_HYBRID_ARTIFACT, produced by the Hybrid Pair Inspector) get applied - and applied to a clone, so the artifact can't corrupt the base's own state. That's the "Use Hybrid -> LoRA ordering" note in the description: artifact first, any LoRA after, never the reverse. If a LoRA lands before the hybrid patch it sees a different tensor layout and the whole Hybrid chain silently misbehaves.

The mode input is your guard rail:

• base_only (default) - stock-loader control. The artifact isn't applied even if you've connected one. Useful as a clean A/B baseline.
• apply_artifact_exp - the experimental path. Requires the exact artifact, and stays flagged experimental because the pack's own testing only claims the chain executes, not that it's a quality win.

The inputs that matter

• quality_base - the exact validated FL2VA-pruned ConvRot checkpoint from your models/diffusion_models folder. The Pair Inspector exists partly so you feed the right file here.
• hybrid_artifact (optional) - the H3_T8_HYBRID_ARTIFACT from the inspector. Leave it disconnected in base_only mode.
• vram_policy (optional) - a T8 VRAM policy applied before the stock load, so the policy is baked into the load path rather than bolted on after.
• weight_dtype - default default; only reach for this if you know the checkpoint needs a specific dtype.

Outputs

model goes downstream to your sampler chain. report_json tells you what was loaded, what mode ran, and whether the artifact actually applied - the difference between "your render looks off" and "your render was off because nothing was applied."

Installing it

Standard pack install: ComfyUI Manager → search "MiniMax H3 Audio T8", or:

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

restart, done. No extra pip packages required and the pack won't download H3 weights for you - the pruned FL2VA/Ref2VA checkpoints are yours to source. If the quality_base dropdown is empty, your checkpoints aren't in models/diffusion_models or ComfyUI hasn't been restarted to see them.

Common issues

The dropdown lists a base but the render is nonsense. Check that the file is actually the FL2VA-pruned base and not the stock 33B checkpoint - the loader trusts the filename but the inspector does not. This is exactly the case the Pair Inspector's SHA verification exists for.

apply_artifact_exp refuses. The artifact isn't the exact one the loader expects. Rebuild it from the Pair Inspector with full_sha256 verification instead of header_only_exp - the header-only mode is explicitly diagnostic and always blocks construction.