(Deno) MiniMax H3 Acc LoRA Loader

MiniMax H3 is a ~33B omni-modal video model, and if you've tried it you know the painful part isn't the prompting - it's how many sampling steps the default schedule wants. Alibaba, the lab that did some of the H3 finetune work, published official acceleration LoRAs that cut that down to 8 steps. The catch: they ship as raw safetensors with a special set of "PDD output heads" bolted on, so you can't just drop them into ComfyUI's normal Load LoRA node. (Deno) MiniMax H3 Acc LoRA Loader is the adapter that makes them work.

Think of it as an LCM-style acceleration for H3, but first-party - Alibaba made the weights for its own model, so you're not gambling on a community conversion. The Deno node loads the official MiniMax-H3-Acc-LoRAs file directly, no converting, no duplicating the safetensors. And it does it with exactly two inputs and one output, which is the whole point of the Deno pack: less setup friction, more actually finishing your render.

How it works

Under the hood the node reads the Acc checkpoint, validates it against the model you connected, then does two things: it applies the static LoRA weights, and it loads the checkpoint's 32 time-dependent PDD output heads. At sampling time a small runtime watches the actual sigma boundaries ComfyUI hands it and fuses those PDD heads for the intervals you're actually running - which is why sampler, scheduler, and step count all stay in the normal ComfyUI nodes instead of being swallowed by the loader.

The interesting bit is how it handles Comfy-Org's curve-pruned H3 checkpoints. A pruned model has a narrower AdaLN layout, so the loader hunts for a matching full checkpoint under models/diffusion_models/, reads only its small FP32 time-embedder section, and builds an in-memory bridge that rebases the 50 full-width AdaLN LoRA updates onto the pruned curve. It never loads the full checkpoint for that. If no full model is installed, it falls back to a compatibility mode: warns you once, skips those 50 updates, and still applies everything else plus the PDD heads. Not ideal, but it doesn't brick your graph.

Inputs and outputs

Only two inputs matter, and they're both required:

• model - a native ComfyUI MiniMax H3 diffusion model. Full and Comfy-Org *_pruned_* variants both work.
• acc_lora - a dropdown of Acc-LoRA .safetensors files. Match FL2VA with FL2VA/T2VA models and Ref2VA with Ref2VA models; the files are picked up from models/loras/ or the dedicated models/minimax_h3_acc_loras/ folder.

The single model output is your patched H3 model - wire it into the normal guider path and build the sampling lane with stock nodes. The recommended starting point is BasicScheduler: simple, steps: 8 → KSamplerSelect: euler → SamplerCustomAdvanced. Keep the native H3 sigma shifts at 12.0 / 3.0 and LoRA strength at 1.0. Simple scheduler will accept 4–12 steps without touching the loader, but treat anything off 8 as an experiment, not an improvement.

Installing it

Grab the pack through ComfyUI Manager (search "Deno Custom Nodes") or clone it manually:

cd ComfyUI/custom_nodes
git clone https://github.com/Deno2026/comfyui-deno-custom-nodes
cd comfyui-deno-custom-nodes
python -m pip install -r requirements.txt

That requirements.txt is just openai-whisper - a dependency of the pack's audio-transcript node, not this one, so the loader itself has no extra baggage. This node needs a recent ComfyUI with native MiniMax H3 support (the pack's H3 nodes require 0.30.0+), so update ComfyUI Stable if the H3 loaders are missing. The Acc weights aren't bundled; grab FL2VA-Acc-8Step.safetensors or Ref2VA-Acc-8Step.safetensors from alibaba-pai/MiniMax-H3-Acc-LoRAs and drop it in one of the two folders above.

Where people get burned

• Empty dropdown. The file isn't in models/loras/ or models/minimax_h3_acc_loras/. That's the whole detection list.
• A warning about a curve-pruned model. You're in compatibility mode. Install a matching full checkpoint so the loader can build the bridge, or accept the skipped updates.
• Old workflows saved with the earlier three-output loader. That version output its own sampler and sigmas. Reconnect them through stock ComfyUI nodes after updating.

One more thing worth knowing before you invest a GPU afternoon: the base H3 weights sit under a Community License that geofences the US, EU, UK and South Korea out entirely - the Acc LoRA is Alibaba's and the loader doesn't change the legal picture for the model it's patching. Outside those regions, though, this is the fastest honest way to render H3 without fighting a conversion script.