VRGDG MiniMax-H3 Turbo LoRA Compatibility

The Turbo LoRA is the speed cheat for MiniMax H3: load it on the base model and you get distilled-style output in a fraction of the steps - the same trick LTX users know from stacking distilled LoRAs. But H3's pruned/fp8 variants come with a nasty surprise: pruning can strip out part of the reference-audio conditioning path, so a Turbo LoRA that works on the full model either errors or silently ignores your audio references on the pruned one. VRGDG MiniMax H3 Turbo LoRA Compatibility is the adapter that makes the Turbo LoRA behave on those pruned checkpoints.

How it works

On the surface it's a LoRA loader: model in, lora_name picked from your loras folder, strength (default 1, range −10 to 10) applied, MODEL out. Underneath, it's doing the compatibility shim - when it detects a pruned model that's missing the reference-audio conditioning pieces, it patches or re-routes them so the Turbo LoRA applies the way it was designed to. That's the whole reason this node exists as a separate class instead of just "load another LoRA."

The strength control deserves more respect than it usually gets with turbo/distilled adapters. A Turbo LoRA is a distilled-speed adapter, and it's the rare LoRA where less is often more - the distillation community's whole playbook is that these adapters change the model's step economy, not just its style, so cranking strength to 2 won't make it faster, it'll just distort. Start at 1, go down if output looks overcooked, and remember negative strengths are available if you're doing something deliberately weird.

The lora_name dropdown pulls from ComfyUI's standard models/loras folder, so your H3 Turbo LoRA goes in the same place every other LoRA lives. If the dropdown is empty, the file isn't there - the pack doesn't ship model weights.

Where it fits

This is the compatibility glue for the H3 branch of the pack's "train standard video, audio, audio/video, and Speed LoRAs" workflow set. If you're running the full fp8 H3 model and want the fast path, this is the loader to reach for instead of the generic one; if you're on the full-size model, the plain loader is fine and you don't strictly need this node - the compatibility layer only earns its keep once pruning is involved.

Install

Pack-wide: ComfyUI Manager → search vrgamedev (or add https://github.com/vrgamegirl19/comfyui-vrgamedevgirl), restart, hard-refresh. Manual: git clone https://github.com/vrgamegirl19/comfyui-vrgamedevgirl into custom_nodes, pip install -r requirements.txt; on Windows portable, install Cython and scikit-build-core first.

Troubleshooting

Empty lora_name dropdown = your Turbo LoRA isn't in models/loras (or ComfyUI hasn't refreshed - check the file's actually there). If the LoRA loads but audio references stop working, that's the exact symptom this node exists to fix - make sure you're using this loader, not the generic one, on the pruned model. And if output quality tanks at strength 1, treat it like any distilled adapter: dial strength down before you start rebuilding the workflow.