Stack_LoRA

Every serious node pack ends up with its own version of this node, because wiring three separate LoRA Loaders in a row is annoying and clutters the graph for no reason. Stack_LoRA is Apt_Preset's take: pick up to three LoRAs and their weights on one node, and if three isn't enough, chain a second Stack_LoRA off the first via lora_stack and keep going. It doesn't apply anything itself - it just builds up a list that gets applied all at once by whatever loader node consumes it downstream, which is the same design rgthree's Power Lora Loader and the old Efficiency Nodes lora stack use, and for the same reason: applying LoRAs one at a time to the model/clip pair is slower and messier to wire than collecting them first and patching once.

The three slots. Each of lora_name_1/2/3 is a dropdown populated from whatever's sitting in your models/loras folder, paired with a weight_1/2/3 float. The weight range here is unusually wide - -10 to 10 - which is more headroom than you'll ever sanely use. In practice, community convention for SDXL-era LoRAs is 0.5-0.8; going much past 1.0 tends to overpower the base model's own aesthetic, and negative weights invert the LoRA's effect rather than just weakening it, which is a neat trick for occasionally undoing something a LoRA does but not a default you want lying around at -1 by accident. Newer architectures like Z-Image run their character LoRAs at 1.0-1.5 as standard, so "correct" weight really is architecture-dependent - check what the LoRA's own page recommends before assuming 0.7 is safe.

Chaining. The lora_stack optional input is what makes this scale past three: feed the output of one Stack_LoRA into the lora_stack input of a second, and the two lists concatenate. This is the same pattern the pack uses for its text and Redux stack nodes too - small fixed-size nodes that compose rather than one giant node with a slider for "how many LoRAs."

Output: lora_stack, of the pack's own LORASTACK type, which downstream Apt_Preset loader/sampler nodes read and apply against the model and clip together.

Installing it. No special dependencies for this one - it's part of the base pack. git clone https://github.com/cardenluo/ComfyUI-Apt_Preset into custom_nodes and restart ComfyUI, or use ComfyUI Manager and search the pack title. Windows users should still run install.bat once, since Manager's dependency resolution can miss packages that other nodes in this pack (not this one specifically) require.

Common gotchas. If a dropdown shows only "None," you don't have any .safetensors LoRA files in models/loras yet - that's not a bug, the enum genuinely has nothing to populate from. Stacking multiple LoRAs and getting a blown-out or NaN-black image is a known interaction hazard with LoRA stacking in general, not specific to this node - try reordering so the largest-effect LoRA goes first, and add them one at a time with a fixed seed to isolate which one is actually causing the problem. And remember a LoRA only works on the architecture it was trained for: an SDXL LoRA plugged into a Flux or Wan pipeline through this node will load without error and simply do nothing, which is a much more confusing failure than an outright crash.