LoRA Notes#
Practical advice for stacking and weighting LoRAs, plus the ones Eimi adds without asking and the one that is currently broken. Start on LoRAs if you have not applied one yet, and use the LoRA catalog for the full list.
Stacking#
Every picker gives you four slots. /settings holds LoRA slots 1-4, the Remix LoRA submenu picks up to 4, and /imagine lora: fills exactly one and blanks the rest for that run.
| Model family | Slots the job can carry |
|---|---|
| Z-Image (Z Image Turbo, Z Image Base, Z Anime, Cyber Realistic Turbo, ZiT-ANI) | No fixed limit. Everything you pick, plus anything auto-applied, is loaded |
| Everything else (SDXL, Anima, Krea 2, Klein, Ideogram 4) | The first 4 only. A fifth is silently dropped |
In practice you set four in the pickers either way. The difference matters when an auto-applied LoRA lands on top of a full stack: on Z-Image it is added, elsewhere it can push a pick past the cut.
Weight guidance#
- Leave the strength blank and you get the LoRA's own tuned default. That default is whatever the LoRA's author or the maintainer put in the config, and is usually the right starting point.
- 0.4 to 1.2 is the useful band. Below 0.3 most LoRAs stop doing anything visible.
- Above about 1.3 a style LoRA starts flattening detail, washing colour and warping hands and faces. Character LoRAs start ignoring your prompt.
- Negative weights push away from what the LoRA learned. This is rarely what you want and often produces mush.
- The LoRA guide walks through choosing a weight for a specific look.
Stacking several
Weights add up in effect, not just individually. Four LoRAs at their 1.0 defaults will usually overcook the image. Drop each to roughly 0.5-0.7 when you run three or four together, then raise the one you care about most.
Two more things to expect when stacking:
- Every selected LoRA's trigger words are injected, so a four-LoRA stack can add a lot of text to your prompt and crowd out what you actually asked for. The character LoRAs have especially long triggers.
- If any LoRA in the stack forces a sampler, that sampler applies to the whole job, and the first one with an override wins. Stack order therefore changes your sampler. The full list of overriding LoRAs is on the LoRAs page.
Known issue: one LoRA cannot be applied#
Realistic Snapshot (Krea 2 v0.5) is unreachable
Realistic Snapshot (Krea 2 v0.5) declares compatibility with a model type called krea. No model in Eimi uses that name -- the three Krea checkpoints (Krea, Krea (Quality), Krea (Exp)) all declare krea2. Because nothing matches, the LoRA is filtered out of every picker: the lora: autocomplete, the /settings LoRA screen and the Remix LoRA submenu. There is no way to select it, and no workaround from the user side. Use Krea Realism or Lenovo UltraReal (v1.0 Krea2) for a similar effect on Krea 2 until the entry is corrected.
LoRAs Eimi applies for you#
Two mechanisms add LoRAs you never picked. Neither appears in your slot list, and neither counts against your four slots as a pick.
Model-attached LoRAs#
Some models are a base checkpoint plus a LoRA, bundled as a single entry in the picker. Choosing the model attaches the LoRA automatically at a fixed strength.
| Model | LoRA it attaches | Strength | Why |
|---|---|---|---|
| Anima Turbo | Anima Turbo | 1.0 | This is the entire difference between it and plain Anima. It is what makes it run at 12 steps and CFG 1 |
| Krea (Exp) | Krea 2 turbo | 0.6 | The experimental Krea checkpoint is not pre-distilled, so the turbo LoRA is stacked on at reduced strength |
The plain Krea model needs no such attachment -- that checkpoint is already distilled.
The NSFW LoRA#
Only when your NSFW setting is on
When NSFW 🔞 is on in /settings, or you pass nsfw: True on /imagine, any LoRA flagged for automatic NSFW injection that is compatible with your model is added for you.
There is exactly one such LoRA:
| LoRA | Applies to | Strength |
|---|---|---|
| S1LV3RC01N Filter Bypass | Any Krea 2 model: Krea, Krea (Quality), Krea (Exp) | 1.2 |
If you had already picked that LoRA yourself, your strength is kept instead. On every model family other than Krea 2, turning NSFW on adds no LoRA at all; it only stops the model's safe-content prompt additions from being applied.
The details panel on the result card shows your own picks, not the injected entry, so a Krea 2 job run with NSFW on will look LoRA-free there even though the bypass ran.