⚗ Alchemica · LoRA (stack of 6) [1.0, use Alchemica LoRA Stack (6 slots + pass-through)]

Six LoRAs means six loader nodes chained model-to-model, each one adding a wire and a place for something to go wrong. This node does the chaining internally: six slots, six strengths, one output.

The slots

lora_1 through lora_6, each with a matching forza_1 to forza_6. The default in each slot is (nessuna) - "none" - so an unused node costs nothing and an unused slot is off by name rather than by strength.

The strength range is the part worth noticing: −4 to +4, step 0.05, default 1. Zero switches a slot off, and a negative strength inverts the LoRA - the author's tooltip says "opposite effect", which is a real tool and not just a curiosity. Some style adapters have a usable negative direction, and it's the fastest way to find out what a LoRA is actually doing to your output: run it at −1 and see what disappears.

clip (optional) - connect the text encoder here and the LoRAs also act on it. forza_clip is a single multiplier for that text-side application, default 1, same −4 to +4 range.

That optional clip socket is the feature. ComfyUI core's LoraLoaderModelOnly deliberately doesn't touch the encoder, and the full LoraLoader wants both a model and a clip and applies the same strength to each. This node lets you say: full strength on the weights, half strength on the text side, or vice versa. On Krea 2 specifically that matters, because...

Krea 2 and LoRAs, briefly

The Krea 2 LoRA scene is the healthiest thing about the model. ai-toolkit had zero-day support and the supply arrived within days - and the community's shared conclusion is that most style LoRAs are unnecessary, because the base model carries so much character and world knowledge already. Where LoRAs genuinely earn their place is identity and subject work, and Krea's own training guidance points at the configuration worth knowing here: train on Raw, apply on Turbo. A community version of the same idea became the default recommendation - Raw plus the official Turbo LoRA at around 0.6 weight, 12 steps instead of 8, euler_cfg_pp with a beta scheduler if you're at CFG 1.

That recipe is exactly what this node is for. Turbo LoRA at 0.6 in slot 1, your character LoRA at 0.8 in slot 2, and your sampler still gets one clean model wire.

There's one silent trap from that same thread and it's worth repeating wherever LoRAs on Krea 2 come up: if you run Raw plus the Turbo LoRA at CFG 1, you need a negative prompt slot wired up even if it's empty. Leaving it disconnected produces grainy output that looks like a sampler problem.

Outputs

model → your sampler. clip → anything that encodes text (the enhancer, the text nodes), and it's only meaningfully patched if you connected a clip input. riepilogo is a STRING summarising which LoRAs were applied at what strength - underrated when you come back to a workflow in a week.

riepilogo is also the sane answer to the general LoRA-stacking problem: no silent skipping, no guessing whether slot 4 was active.

Wiring, including the hybrid case

⚗ Load Krea 2 · model → ⚗ LoRA (stack of 6) · model → AlchemicaKrea · model

And the reason the pack splits composition from refinement: on the hybrid path you can give the composition model one stack and the refinement model a different one, so a style LoRA shapes the picture while an anatomy finetune handles the finish. That's what AlchemicaKrea's modello_rifinitura socket and the pack's KREA2_IBRIDO example exist for.

Installing

Part of the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/Nessuno76/ComfyUI-AlchemicaSampler

Restart the ComfyUI server. New slots appear in the menu on startup, so if you just dropped a LoRA into models/loras and it isn't listed, that restart is the fix. No dependencies; ComfyUI 0.36.0+.

Traps

LoRAs from other architectures. A LoRA trained for SDXL or Flux will load and produce noise. Krea 2 is a 12B single-stream DiT with a Qwen3-VL encoder; nothing else is compatible. Also: FAL-format LoRAs need converting before ComfyUI will load them properly.

Stacking five and wondering why it's mush. Each LoRA is a foreign influence on a distilled model. Four or five at full strength is where the community reports texture errors and grain, and the standard fix is more steps - one user's note was that with five LoRAs, 12 steps instead of 8 cleared the artefacts.

Setting forza_clip and not connecting clip. It does nothing; that output stays unpatched.

Expecting the box's (nessuna) to mean 0 weight. It means the slot is not applied at all. A LoRA at strength 0 is loaded and multiplied by zero, which is not free.