LoRAs#
A LoRA is a small add-on file that pushes a model toward a style, an artist, a character or a concept without changing the model you picked. There are 91 LoRAs loaded at the moment, all of them enabled and reachable. The full list is on the LoRA catalog; practical advice and known problems are on LoRA notes. For a walkthrough of choosing and blending them, see the LoRA guide.
The library is just files on the maintainer's drive. New ones get added and old ones get deleted whenever they feel like it, with no announcement, so a LoRA you used last week can simply be missing. The picker in Discord is the only accurate list.
How to apply one#
There are three places a LoRA can come from. They all end up in the same four slots on the job.
| Route | Where | What it gives you |
|---|---|---|
lora: parameter |
/imagine |
One LoRA for that single run, with an optional lora_strength: |
| LoRA slots 1-4 | /settings -> Models & Tools 🎨 |
Up to four LoRAs applied to every /imagine run, each with its own strength and variant |
| LoRA submenu | The Remix menu on a result card, Configure ✨ | Up to four LoRAs for the re-run you are about to launch |
/imagine prompt: a knight in a rainstorm lora: Neurocore lora_strength: 0.8
lora: replaces your saved stack
Passing lora: on the command blanks saved slots 2, 3 and 4 for that run, and ignores the variant you saved for that LoRA. Leave lora: out if you want your saved four.
Remix does not auto-load your saved slots, because the model Remix suggests may not be the one your slots were chosen for. Pick LoRAs in the Remix LoRA submenu if you want any.
Compatibility#
Every LoRA carries a list of compatible model types. A model qualifies when its model type, or any of its subtypes, appears in that list. An empty list would mean the LoRA works with every model.
No universal LoRAs right now
Every one of the 91 LoRAs currently names at least one model type, so there is no LoRA that works everywhere.
This is the whole mapping. Pick your model first, find its row, and you know which LoRA group to shop in.
| Compatibility token | LoRAs | Models it unlocks |
|---|---|---|
z-image |
44 | Z Image Turbo, Z Image Base, Z Anime Base, Z Anime 8-Step, Z Anime 4-Step, Cyber Realistic Turbo, ZiT-ANI |
sdxl |
28 | Every SDXL checkpoint: mystery i32, FantasyAnime (v2.0), WAI (v16.0), One Obsession (v18), Cat Tower (v20 v-pred), Cat Carrier (v9.0), Echo Dream (v1.0), Cyber Realistic Pony (v16), Milk (Walnut), Hassaku (v3.0), Dreambox (v4.0), Visionary (Surreal Chaos), Red Lily (Artsy), JANKU (v6.9), Noob AI XL (vPred 1.0), Nova Anime XL (v8), Quillworks (V2), RedLilly (v1.0), waiNSFW (v13.0), WAI-SHUFFLE (v-pred-04) |
anima |
7 | Anima, Anima (aesthetic-v1.1), Anima Base (v1.0), Anima Cat Tower (v1.0), Anima Turbo, AnimaYume (v0.5), WAI Anima (v1.0) |
cosmos |
1 | The same Anima models except Anima (aesthetic-v1.1) |
krea2 |
5 | Krea, Krea (Quality), Krea (Exp) |
klein_9b |
3 | Klein 9B Fast, Klein 9B Quality |
klein_base |
1 | Klein 4B Quality, Klein 9B Quality |
ideogram4 |
2 | Ideogram 4, Ideogram 4 (Fast), Ideogram 4 (Quality), Ideogram 4 (Turbo) |
krea |
1 | Nothing. See the warning on LoRA notes |
Some models have no compatible LoRAs at all: Flux Dev, Qwen, ERNIE Image, ERNIE Turbo, Mage Flow (Quality), Mage Flow Turbo, Mugen Aesthetic, NetaYume AIO (v3.5) and Klein 4B Fast. Their pickers will say there is nothing compatible.
You never have to work this out by hand in normal use. Once a model is set, every picker (the lora: autocomplete, the /settings LoRA screen, and the Remix LoRA submenu) filters itself down to compatible entries only. Working the other way round also holds: if you pick a lora: first on /imagine, the model: list narrows to models that LoRA fits.
See Models for what each model type actually is.
Weights#
Each LoRA has a tuned default strength stored with it. Leave the strength field blank and you get that default, or 0.8 if the LoRA has none. Defaults across the catalog run from 0.4 to 2.0; most sit at 1.0.
| Where | Range you can set |
|---|---|
/imagine lora_strength: |
-10 to 10 |
/settings LoRA strength, per slot |
-10.0 to 10.0 |
| Remix Edit LoRA Weights | -2.0 to 2.0 per slot |
The useful band is roughly 0.4 to 1.2. Negative values push away from what the LoRA was trained on. A few entries are tuned deliberately low (0.4-0.5) because they are strong, and a couple are tuned above 1.0 because they are weak; the per-LoRA defaults in the catalog are the ones to trust before you start experimenting.
Trigger words#
Most LoRAs were trained with an activation phrase. Eimi injects it for you when the LoRA is selected, so you write your prompt normally and never type the trigger yourself. Triggers are appended to your positive prompt, except for Dino (v1.0), whose triggers go to the front of the prompt instead.
24 of the 91 LoRAs have no trigger at all and work on selection alone.
raw mode drops triggers
Running with raw: True on /imagine, or the RAW 🔧 flag in /settings, sends only your text to the model. LoRA trigger words, model quality tags and embeddings are all dropped. The LoRA file is still loaded, but nothing activates it.
Variants#
Four LoRAs are packs that hold more than one character or outfit. Picking a variant swaps in a different trigger phrase from the same file. Variants appear in the pickers as separate rows reading Nickname (Variant name).
| LoRA | Base entry | Variants |
|---|---|---|
| Momo Ayase (Ill) | Momo Ayase | School Uniform, Casual, Gym Clothes |
| Amity Blight - S3 (Ill) | Amity Blight S3 | Overalls, Hecate Costume, Boo Outfit, Finale Outfit |
| The Owl House Style Pack (v4.0 Ill) | Style only | Luz Noceda, Amity Blight, Eda Clawthorne |
| Aubrey (Omori) | Headspace | Far Away |
In /settings you can also set a variant per slot after the fact, through the Variants 🎠button that appears once a slot holds a variant-carrying LoRA.
LoRAs that change your sampler#
20 LoRAs carry their own sampler, or sampler and scheduler pair, because they were trained with it. Selecting one overrides the sampler and scheduler for the whole job, including the model's own defaults. If you stack several, the first one with an override wins. Passing sampler: or scheduler: yourself does not survive this.
| LoRA | Sampler | Scheduler |
|---|---|---|
| Bright Oni (v1.0 Z-Img) | euler |
model default |
| ta86 style (v1.0 Z-Img) | res_multistep |
simple |
| Neurocore (Z-Img) | euler_ancestral |
simple |
| Lenovo UltraReal (v1.0 Z-Img) | euler |
sgm_uniform |
| Illustrious Style (v2 Z-Img) | er_sde |
simple |
| Simon Stalenhag Style (Z-Img) | dpmpp_2m_sde |
sgm_uniform |
| Butternut Crunch Wilds (v1 Z-Img) | euler_ancestral |
beta |
| Transparent Watercolor (v2 Z-Img) | res_multistep |
simple |
| PSX Style (Z-Img) | euler |
simple |
| Anime Lines (Z-Img) | res_multistep |
model default |
| Retro Neon Cyberpunk (Z-Img) | euler |
simple |
| Pixel 6x6 (Z-Img) | euler |
simple |
| Line Melts (v2 Z-Img) | uni_pc |
sgm_uniform |
| YKK Style (v1 Z-Img) | dpmpp_sde |
model default |
| Flat Color (v2.1 Z-Img) | euler |
simple |
| Modern Anime Render (v1 Ill) | euler_ancestral |
normal |
| Neurocore (Ill) | euler_ancestral |
normal |
| Musaigen Custom (v1 Anima) | euler |
sgm_uniform |
| Krea Realism | er_sde |
sgm_uniform |
| Lenovo (Ideogram 4) | res_2m |
model default |
The details panel on the result card shows what actually ran, so you can check which sampler you ended up with.
Tags stand in for categories#
There is no formal category field. Each LoRA carries a list of tags, and those tags are what the search scores against, so they behave as the category system.
| Tag | LoRAs | What it groups |
|---|---|---|
style |
74 | Anything that changes how the image looks rather than what is in it. Nearly everything |
anime |
41 | Anime and illustration styles |
retro |
13 | VHS, PSX, film, 2000s and 90s looks |
realistic |
12 | Photographic and semi-realistic looks |
character |
11 | Named characters from a series |
photography |
7 | Camera-imitating LoRAs: film grain, flash, phone cameras |
illustration |
7 | Painterly and illustrative rendering |
dark |
6 | Low-key, moody, high-contrast looks |
artist |
5 | Imitations of one named artist |
cartoon |
5 | Western cartoon styles |
sketch, charcoal, watercolor, traditional |
3-4 each | Traditional media |
pixel_art, psx, 3d, gaming, low_poly |
2-3 each | Game-adjacent and pixel looks |
flat |
4 | Flat colour, no lineart |
detailed, painting, 2000s |
4 each | Detail-heavy, painterly and 2000s-era looks |
owl_house, omori, crosscode, dandadan, arknights, ruri_dragon, wonder_egg_priority |
1-4 each | Source series for the character LoRAs |
turbo, speed, distill |
1-2 each | Step-reduction LoRAs |
meme, food |
2 each | The two joke entries |
nsfw |
2 | Adult-oriented entries |
Searching the picker#
The picker does not browse alphabetically, it searches. Type a fragment and results come back scored, so partial and approximate names work:
- An exact start-of-name match ranks highest, then a word starting with your text, then a match anywhere inside the name.
- Variant names, tags and words from the description are also matched, so you can search by concept (
watercolor,cyberpunk,owl house) rather than by filename. - Ties break on the LoRA's display priority.
- With the search box empty you get the list in priority order, and the first entry that has variants gets its variants shown inline.
LoRAs you have hidden in /settings are filtered out after scoring and never appear.