MiniMax H3 LoRA Loader

Character consistency, a recurring style, a stable location - in the ComfyUI world that's what LoRAs are for, and H3 is no exception. The catch is that H3's DiT is a pruned, distilled model, and the LoRA ecosystem around it has its own rules. This loader is built for one specific format - the "complete pruned LoRA" produced by the author's companion LoRA tool - and knowing that up front saves you an afternoon of silent failures.

How it works

An H3 LoRA isn't just backbone weights. To steer the model it also needs the AdaLN (adaptive layer norm) modulation tables - adaln_t_table, plus the blocks.*.adaln_proj and final_layer.adaln_proj linear weights and biases. A "complete pruned LoRA" bundles all of that together, so the loader can apply the AdaLN replacement at model-load time and you never need the separate h3_silu_temb_grid.safetensors file.

The README is explicit about the format gate: original 2688-dim AdaLN Turbo LoRAs are not supported on pruned models. If you have one, you bake it into a complete pruned LoRA first (via the author's tool at xiaolibai-sys/MiniMax-H3-Pruned-Lora-Adapter) and then load that. This is the single biggest gotcha in the pack - the loader isn't broken, the format genuinely doesn't match.

The inputs that matter

• model (required) - a MINIMAX_H3_MODEL from the streaming Model Loader. This node is a pass-through: it takes the model, applies the LoRA, hands back a modified model.
• lora_name - dropdown of LoRA files in ComfyUI/models/loras/.
• strength - default 1.0, range −10 to 10. Below 1 weakens the effect; negative inverts it. For a first pass, leave it at 1.
• silu_grid - Auto (default), FL2VA, or REF2VA. This is the grid the Turbo runtime AdaLN uses. Auto picks REF2VA when the model filename contains "ref2va" - so only touch it if Auto's guess is wrong.

One output: model, the patched MINIMAX_H3_MODEL, wired into the KSampler (chained after the base Loader).

Installing it

Same pack routine - ComfyUI Manager search "MiniMax H3", or:

cd ComfyUI/custom_nodes
git clone https://github.com/xiaolibai-sys/ComfyUI-MiniMaxH3
pip install -r requirements.txt

then restart. The LoRA files themselves go in ComfyUI/models/loras/. And the LoRA training side is a separate repository (MiniMax-H3-Pruned-Lora-Adapter) - this node only loads.

Common issues

• LoRA loads but does nothing / errors at load. Almost always the format. Check whether it's a 2688-dim Turbo LoRA; if so, it needs baking into complete-pruned form first. The README calls this out as the unsupported case.
• LoRA not in dropdown. Files land in models/loras/ before startup; refresh after dropping one in.
• silu_grid mismatch. If a REF2VA-trained LoRA is running on a model whose name doesn't say "ref2va," Auto may guess wrong. Set it manually to REF2VA.