加载LoRA（旁路，仅模型）（用于调试）

The display name is Chinese - 加载LoRA（旁路，仅模型）（用于调试）- and it translates to "Load LoRA (bypass, model only) (for debugging)." That long parenthetical is the honest description of what this node does and doesn't do. It doesn't load a model into your sampling chain, because there is no sampling chain here. It builds a FEIHOU_MERGE_LORA_STACK - a list of LoRA adapters with strengths and optional bypass flags - and hands that stack to the pack's Loader, which applies the LoRAs to the H3 transformer internally. "Bypass, model only" means the LoRA data is passed along the wire to whoever owns the model, never inserted into a downstream MODEL chain itself.

Why it exists

In the vanilla ComfyUI pattern you'd chain LoRA loading onto a model with the LoRA node's model output feeding onward. This pack deliberately doesn't do that - the H3 transformers live inside the loader, so a LoRA node that wants a MODEL in one pin and out the other would have nowhere to plug in. The stack node sidesteps the whole problem: it produces only the stack, the loader consumes only the stack, and the model never moves. The "(for debugging)" tag is the author admitting this is the low-level tool you use when the fancier path misbehaves - if you're chasing which of six LoRAs broke your character, being able to flip each entry on and off without rebuilding the model chain is exactly what you want.

How it works

Mechanically it's nearly trivial. The stack function takes an optional incoming optional_lora_stack (so you can chain stacks together), normalizes it to a list, then appends every lora_* input it received - the node grows a new input slot for each LoRA you add, in the same native canvas-stack style as the author's FeiHou LoRA Stack (Merge/Extract). Output is a single lora_stack.

Wiring it up

Place it to the left of the Loader, connect lora_stack → the Loader's left-side LoRA stack input. The Loader applies every enabled entry and packs the result into the bundle. If you're using the Remix Loader, you get two stack inputs (first_pass_lora_stack and second_pass_lora_stack) and can feed separate stacks to each pass. Not using the pack's loaders at all? Then skip this node - the Model Adapter path accepts standard MODEL wires and has no stack input.

How to install it

Pack install, same as its siblings:

cd ComfyUI/custom_nodes
git clone https://github.com/FX-FeiHou/ComfyUI-FeiHou-Easy-H3

restart, or ComfyUI Manager → search "ComfyUI-FeiHou-Easy-H3". No pip dependencies, no models of its own - LoRA files live in your normal models/loras folder and show up as you add entries.

Common issues

The one trap worth naming: because this node produces no model and touches no sampler, it's easy to connect it and expect results, then wonder why nothing changed. If the stack isn't visibly doing anything, check that it actually feeds the Loader's lora_stack input (not the main node directly), that entries are enabled rather than bypassed, and that your LoRA was trained for H3's architecture - a FLUX LoRA on a MiniMax H3 transformer is just noise, exactly the architecture-lock-in reality every modern LoRA carries.