LoRAs#
A LoRA is a small add-on file that shifts a model toward a character, an outfit, a style or a concept. This page covers how the stack behaves. For the list of what is actually installed, see the LoRA catalog.
The stack#
LoRAs are applied on top of the checkpoint you picked, in the order they sit in your slots. You get 4 slots.
| Family | Slots | What the LoRA affects |
|---|---|---|
| SDXL / Anima / Mugen / Flux 2 / ERNIE / Qwen (bundled) | 4 | model and text encoder |
| Z-Image / Lumina2 | unlimited | model and text encoder |
| Krea 2 | 4 | model only |
| Qwen Nunchaku | 4 | model only |
| Ideogram 4 | 4 | model only, applied to both of its branches |
/turbo TensorRT presets |
0 | LoRAs are not supported at all |
/turbo MrFlow on Z-Image or Krea |
4 | model and text encoder |
/turbo MrFlow on Qwen or Ideogram |
0 | any LoRA you pass is silently dropped |
Z-Image is the only family with no cap, because it uses a different loader node.
Warning
On a /turbo MrFlow Qwen or Ideogram run, LoRAs are dropped without an error message. If your style did not apply on a turbo job, this is why.
Compatibility#
Every LoRA declares which model types it works with. The rule is short:
- A LoRA that declares no compatible types is universal and is offered on every model.
- Otherwise it is offered only when the model's type or any of its subtypes appears in that list.
Subtypes matter. An SDXL checkpoint may carry subtypes such as illustrious, noob, pony or vpred, and a LoRA listing any one of those will show up for it.
The pickers do the filtering for you. Once you have selected a model in /settings, the LoRA list is narrowed to compatible entries only. The /turbo picker resolves a turbo preset back to its underlying base model before filtering, so model-specific LoRAs still appear there.
The library holds 91 entries at the moment. Compatibility spread: Z-Image 44, SDXL 28, Anima 7, Krea 2 5, Klein 9B 3, Ideogram 4 2, Klein base 1, Cosmos 1, Krea 1. Entries come and go as the maintainer downloads and prunes files, so that count is a snapshot rather than a fixed size.
Weights#
| Default | the LoRA's own configured default, or 0.8 if it has none |
Range on /imagine lora_strength: |
-10 to 10 |
| Range in the settings strength modal | -10.0 to 10.0 |
Configured defaults in the installed library run from 0.4 to 2.0, so "default" is not always 0.8. Above roughly 1.0 most LoRAs start overwhelming the checkpoint; negative values push away from the concept.
lora_strength: on /imagine applies to the single inline LoRA only. Per-slot strengths for your saved LoRAs are set with the Strengths ⚙️ button in the LoRA picker.
Trigger words#
Most LoRAs need a trigger word in the prompt to activate properly. You do not have to type it -- the bot adds it for you.
- The configured trigger text is added to your positive prompt, and some LoRAs also add text to the negative prompt.
- Placement is per-LoRA: a trigger is either prepended (the default) or appended.
- When AI enhancement is on and the LoRA carries a placement hint, the local language model is asked to weave the trigger into the sentence at a sensible spot instead of bolting it on the front. Character triggers land near the character, style triggers at the front, outfit triggers right after the character trigger. This call has a 3-second timeout; if it does not finish, the plain prepend happens instead.
Variants#
Some LoRAs ship multiple variants -- alternate trigger sets for different characters or outfits packed into one file. Examples in the library: Amity Blight with overalls / hecate / boo / finale, The Owl House with luz / amity / eda, Momo Ayase with outfit_1 through outfit_3.
Variants replace, they do not add
Selecting a variant replaces the base trigger entirely. Variants are self-contained tag sets, not extras layered on top of the default trigger.
In the LoRA picker, variants appear as their own rows, and a LoRA with an active variant floats to the top of page 1. The Variants 🎭 button opens a per-LoRA dropdown for setting your default variant. That button only appears if at least one of your active LoRAs actually has variants.
Sampler and scheduler overrides#
A LoRA can declare its own sampler and scheduler. The first LoRA in your stack that declares one wins, and it overrides the model's default. This is why picking certain style LoRAs silently changes the sampler shown on your result card.
Priority, highest first:
- An explicit
sampler:/scheduler:you passed on the command - The first LoRA in the stack that declares one
- The model's own default
See Parameters for what the samplers themselves do.
The settings slots#
Open /settings → LoRAs to fill your four slots. The picker is a multi-select, 20-ish entries per page, filtered to your selected model.
| Control | Emoji | What it does |
|---|---|---|
| Strengths | ⚙️ | Opens a modal with one field per active LoRA |
| Variants | 🎭 | Per-LoRA variant dropdowns, only shown when an active LoRA has variants |
| Clear All | 🗑️ | Empties all four slots, their strengths and their variants |
Clear All has no confirmation
It takes effect the moment you click it.
Selection behaviour worth knowing: picks on the page you are looking at replace that page's LoRAs, while active LoRAs on other pages are kept. The result is deduplicated to one slot per file and truncated to 4.
Passing a LoRA on the command#
/imagine prompt: a knight in a storm lora: <pick one> lora_strength: 0.9
lora: throws away your saved slots
Passing /imagine lora: uses that one LoRA and blanks slots 2 through 4 for that run. It also ignores the variant default you saved for that LoRA. If you want your saved stack, leave lora: empty.
Some models also inject LoRAs on their own. Anima Turbo and Krea (Exp) are preset entries that silently load their own acceleration LoRA at a fixed strength. Those do not consume your slots and are not shown in the parameter summary.
Related#
- LoRA catalog -- every installed LoRA, its triggers and its variants
- LoRA overview -- what LoRAs are and how they are organised
- Notes and issues -- known problems with specific entries
- Parameters -- the sampler and scheduler a LoRA can override
- AI Enhance -- the step that places trigger words intelligently