H3 LoRA Stack (up to 4)

If you've ever stacked four LoraLoader nodes in a row just to get one model with a style plus a character plus a quality boost, you know the pain this removes. H3LoRAStack takes a model and up to four LoRAs with per-LoRA strengths, applies them in order, and hands you a single model out the other side. That's the whole job, and it's a genuinely useful one in a long-form video graph, where every node you remove from the path is a place you can't miswire.

How it works

In goes model, out comes model - the same model with the LoRAs merged in sequence. Four lora_N dropdowns enumerate everything in models/loras (plus a None option), and four strength_N floats (-2 to 2, default 1.0) set each one's weight. Order matters, as it always does with LoRA stacking: applied left to right, so if two LoRAs fight over the same concept, whichever is in the lower slot (lora_1) merges first. Negative strengths are allowed - useful for removing a baked-in tendency with a negative-weight LoRA, the same trick you'd use with a textural inversion.

The mechanics are the standard ComfyUI LoRA patch under the hood - it's a convenience node, not a new training method. What it buys you is a single point of control. You can drive the whole stack from one widget panel instead of five chained loaders, and when a LoRA's model produces garbage, you can blame the stack order instead of hunting through a daisy chain.

When you'd actually reach for it

H3's identity system leans on reference photos rather than LoRAs - the cartridge loader's report explicitly notes that LoRA entries are ignored because "H3 identity comes from references, not LoRAs." So don't expect this to replace ref-based character work. Where it shines is style, motion, and quality LoRAs: a camera-motion LoRA, a film-style LoRA, a detail booster, stacked and balanced against each other. One strength too hot? Dial it down in the same panel instead of rewiring.

One thing worth knowing: if a stack renders as though one LoRA isn't there, check whether it's a LoRA trained for a different model family. This node doesn't validate compatibility - it merges what you give it, and an incompatible rank shape fails at load or silently does nothing useful. Same rule as everywhere: only stack LoRAs that were trained for the model you're running, ideally on the same base.

Install

It's one node in the H3 Multishot pack:

cd ComfyUI/custom_nodes
git clone https://github.com/jlucasmcrell/ComfyUI-H3-Multishot

or ComfyUI-Manager → H3 Multishot. Restart; needs ComfyUI v0.30.0+ for the H3 nodes around it. No extra dependencies - the pack's requirements list is deliberately empty.