Eric Krea2 Multi-LoRA Stack

This is the pack's answer to "I need three LoRAs and they don't all matter at the same time." Eric Krea2 Multi-LoRA Stack is a single growable node holding up to ten LoRA slots, and each slot carries its own three weights - one for Stage 1, Stage 2, Stage 3 of a multi-stage run. Apply LoRA is the one-LoRA version of the same idea; this is the version for when a style LoRA should own composition, a face LoRA should carry the refine pass, and a detail LoRA should only touch the final stage.

The mechanics are simple: pick a file per slot from models/loras/Krea2/, and the node fuses the enabled adapters onto the pipeline in slot order. Unused rows auto-hide until you need them, which keeps the panel from being a wall of widgets. Each slot gets:

• on_N - enable/disable without clearing the slot (distinct from none, which is an empty slot).
• lora_N - the file dropdown.
• strength_Ns1 / s2 / s3 - the per-stage weights. Set all three equal for a uniform strength; spread them for stage-scoped control.

Ten slots is a lot. If you're stacking more than four or five, the honest question is whether your LoRAs are doing real work or fighting each other - but the per-stage spread is precisely what makes layered stacks behave, since each adapter can own the stage it's actually good for instead of applying flat everywhere.

Shared controls worth knowing:

• ephemeral (default ON) - load fresh, unload after the run. Off leaves everything resident and forces manual unloading; the tooltip's "if ANY LoRA in the stack is ephemeral, the whole stack is cleared after the run" is the reminder that this is all-or-nothing.
• stack_enabled - master switch for the whole node; pass the pipeline through untouched without clearing your slots.
• lora_preset - ★ Save preset dropdown, storing slots, strengths, and toggles. The shipped samples (Split_Distil, turbo-distill-128-split) hint at the split LoRA tricks people run on Turbo - the community's Raw + Turbo LoRA at 0.6 path is a one-preset recipe.
• fetch_triggers / add_triggers / force_refetch - trigger words for every enabled slot, combined and de-duplicated in slot order into the trigger_words output, optionally merged into the prompt passthrough. Only the first lookup per file touches the network (cached in data/lora_triggers.json).
• prompt - passthrough so merged triggers reach your prompt output.

Outputs mirror Apply LoRA: pipeline, settings, trigger_words, prompt. Wire pipeline into Generate or Multi-Stage Ultra; fold settings into Merge Settings for the reproducibility chain.

Where people get burned: per-stage weights are easy to forget once you have ten slots of them - a slot left at 1.0 across all three stages will apply its full strength at every resolution, and a heavy style LoRA at Stage 3 is a common source of that "looks great at S1, ruined at 4k" complaint. And with ephemeral off, orphaned weights silently follow you between workflows - the Unload node exists for exactly that.

How to install

Part of EricRollei/Krea2_ComfyUI_Advanced - ComfyUI Manager → search "Krea2_ComfyUI_Advanced", or:

cd ComfyUI/custom_nodes
git clone https://github.com/EricRollei/Krea2_ComfyUI_Advanced

Restart. Pack-wide requirements: git-installed diffusers with Krea2Pipeline, transformers>=4.57.0, Krea 2 weights, LoRAs in models/loras/Krea2/.

Common issues

• Slots disappearing - unused rows auto-hide; pick a file to activate a slot, and a new empty one appears below.
• Weights bleeding between runs - ephemeral off. Flip it on or unload manually.
• Trigger words never appear - fetch_triggers must be on, and the node only queries Civitai for files without safetensors metadata; force_refetch is there for stale lookups.
• Stacked LoRAs muddy each other - spread them per-stage instead of stacking all at 1.0 everywhere.