Skip to content

/settings#

Your personal defaults for every other command. /settings opens a private, interactive menu tree where you pick a model, LoRAs, an upscaler, prompts, output format, and more. Everything you set applies to you across every server, DM, and channel.


Synopsis#

/settings

No parameters. The reply is always ephemeral (only you can see it) and works in servers, DMs, and private channels, with either install type.

If you get an access-denied message instead of the menu, see Access Control.


How the menu works#

Behaviour Detail
Auto-save There is no Save button. Every dropdown pick and every toggle writes to the database the moment you press it, then the page redraws.
Green vs grey A green button means you have customized that thing. Grey means it is still at the system default.
Navigation Every page has Back ◀ (previous page) and Home 🏠 (jumps to the top of the tree).
Timeout 15 minutes. After that the buttons stop responding and you re-run /settings.
Ownership Only the person who ran the command can press the buttons.
Takes effect immediately The next command you run already uses the new value.

Note

Undoing is not symmetrical. Reset All asks for confirmation, but Clear All (LoRAs) and Reset Order (custom sort) apply instantly with no confirmation.


Home#

The top of the menu shows a single status line listing only the values you have changed, for example Model: X · LoRAs: 2 · Ratio: 16:9 · RAW · WebP: on · AI: both. If you have never changed anything it reads All settings at defaults.

Button Emoji What lives inside
Models & Tools 🎨 Imagine model, LoRAs, upscaler, edit model, restyle style and model
Generation Aspect ratio, embeddings, default prompts, RAW, NSFW
Output 🖼️ Image format and quality, prompt display, previews, audio
Features 🔬 AI features, match sensitivity, history page size
Describe 🔎 Backend, output style, descriptions per run
Reset All 🔄 Wipe every setting (red button)

Note

Setting only a Restyle style preset or a Restyle Model leaves the Models & Tools button grey. That is cosmetic only -- the setting is saved.


Models & Tools#

Six rows, each with a Change button that opens a picker.

Row Shows Opens
Imagine: model nickname, or Default Model picker
LoRAs: N selected, or None LoRA picker
Upscaler: upscaler nickname and Standard or USDU Upscaler picker
Edit: edit-model nickname, or Default Edit model picker
Restyle: style preset label, or Default Style preset picker
Restyle Model: edit-model nickname, or Same as Edit Restyle model picker

Below them are two read-only buttons: Model Info 📋 (catalog of the models you can pick, 20 per page, with license, type, description, and links) and Internal Models 🔧 (the models Eimi uses behind the scenes: CLIP, VAE, detection, segmentation, background removal, embeddings, enhancement, runtime engines, and upscale methods).

The settings on this page#

Setting Values Default What it does
Imagine model Any enabled model, minus any you have hidden None (Eimi picks its own default) The model /imagine uses when you do not pass model:
LoRA slots 1-4 Any enabled LoRA compatible with your model Empty Up to 4 LoRAs applied to every /imagine run
LoRA strength (per slot) -10.0 to 10.0 The LoRA's own default strength, or 0.8 How strongly each slot is applied
LoRA variant (per slot) Any variant defined for that LoRA None (base trigger only) Swaps in a variant trigger for that LoRA
Upscaler Any enabled upscaler, plus SeedVR2 4x-AnimeSharp.pth Used by /imagine upscaler: and the ⬆️ Upgrade button
Upscale method Standard or USDU Standard Plain model upscaling vs AI-enhanced tiled diffusion (slower, higher quality)
Edit model See the list below Klein 4B Fast Default model for /edit and the 🎨 Edit button
Restyle model Same list as edit model Falls back to your edit model, then Klein 4B Fast Model used for the restyle context menu
Restyle style preset One of the 25 style presets None Default preset for restyle

Edit and restyle model options: Klein 4B Fast, Klein 4B Quality, Klein 9B Fast, Klein 9B Quality, Krea 2 Edit Fast, Krea 2 Edit Quality, Mage Flow Edit Fast, Mage Flow Edit Quality, Qwen Edit Fast, Qwen Edit Quality, Nano Banana, Nano Banana 2, Nano Banana 2 Lite, Nano Banana Pro, GPT Image 2.

VIP only

Nano Banana, Nano Banana 2, Nano Banana 2 Lite, Nano Banana Pro, and GPT Image 2 are only offered to VIP members and admins. See Access Control.

Note

Finishing a restyle silently saves the preset you just used as your new default Restyle style. If your default keeps changing on its own, that is why.


The shared picker#

All six pickers use the same layout.

Element Detail
Header The current value, or System Default, plus Page n/N
Dropdown 24 items per page. The first option is always Clear (Default) 🔄, which resets that setting
Pagination ◀ / n/N / ▶ (only shown when there is more than one page)
Priority 📊 Sort the list by Eimi's configured priority order (the default)
A-Z 🔤 Sort the list alphabetically
Custom ✏️ Open the Custom Order editor
Info 📋 Open the read-only catalog for that category

Picking an option saves it and keeps you on the same page. The sort you choose is saved per category, so models, LoRAs, upscalers, edit models, restyle models, and styles each remember their own sort.

Your sort order and hidden items also drive the model:, lora:, and upscaler: autocomplete lists on /imagine.

LoRA picker extras#

The LoRA picker is a multi-select, so you can tick up to 4 entries at once.

  • The header shows an Active: line with each LoRA, its strength, and its variant.
  • If you have a model set, only LoRAs compatible with that model are listed.
  • LoRAs that have variants get one row per variant, so you can pick a variant directly. Anything currently active floats to the top of page 1.
  • Picks on the current page replace that page's LoRAs. LoRAs you selected on other pages stay active. One slot per LoRA file, four slots maximum.

Three extra buttons appear once at least one LoRA is active:

Button Emoji What it does
Strengths ⚙️ Modal to type a strength for each active slot
Variants 🎭 One dropdown per active LoRA that has variants, plus a Default (base trigger only) option. Only shown when an active LoRA actually has variants
Clear All 🗑️ Empties all four slots, strengths, and variants with no confirmation

Upscaler picker extras#

The upscaler picker puts a Method section above the list, showing Standard | USDU with a Toggle button. USDU is AI-enhanced tiled diffusion upscaling: slower, higher quality. See Upscalers.

Warning

/upscale does not read your saved upscaler. It falls back to the bot-wide default upscaler unless you pass upscaler: yourself. The ⬆️ Upgrade button on a result card does honour your saved upscaler and method.

Custom order and hiding#

The Custom ✏️ button opens an editor for that category, 24 items per page.

Control Emoji What it does
Select item to move -- Pick the row you want to act on. It is marked ▶️ in the list
Up / Down ⬆️ / ⬇️ Move the selected row. Disabled at the ends of the list
Hide / Show 🚫 / 👁️ Hide the item from your lists. Hidden rows are marked 🚫
Reset Order 🔄 Clears your custom order and your hidden items for that category, with no confirmation

Hiding saves instantly. Reordering only saves when you press Back or Home, and only if you actually moved something. Saving a new order also switches that category's sort method to Custom.

Note

Hiding is cosmetic. A hidden model or LoRA is removed from your pickers and autocomplete, but still works if you type its name into a command manually.

Warning

On the Restyle Model picker, the Custom ✏️ button reports "No items available to sort" and the Info 📋 button shows "No items". Sorting and the catalog are not wired up for that category yet. Picking a restyle model from the dropdown works normally.


Generation#

Setting Values Default What it does
Aspect ratio 1:1, 3:4, 4:3, 9:16, 16:9, 2:3, 3:2, or a custom value None (the model's own default, falling back to 3:4) Output shape for /imagine
Positive embeddings lazypos None Embeddings appended to your positive prompt
Negative embeddings lazyneg, deep_negative_pony, lazynsfw, smooth_negative None Embeddings appended to your negative prompt
Default positive prompt Free text, up to 2000 characters Empty Text folded into every positive prompt
Default negative prompt Free text, up to 2000 characters Empty Text used as your negative prompt when you do not pass negative:
RAW 🔧 On / Off Off Sends only your text. No model quality tags, no embeddings, no LoRA triggers, no SFW additions, and no AI enhancement
NSFW 🔞 On / Off Off Off injects the model's SFW additions. On skips them and auto-applies any LoRA the model flags for NSFW

The aspect ratio dropdown (placeholder 📐 Aspect Ratio) offers fixed pixel sizes:

Option Pixels
1:1 Square 1024 x 1024
3:4 Portrait 896 x 1152
4:3 Landscape 1152 x 896
9:16 Portrait 768 x 1344
16:9 Landscape 1344 x 768
2:3 Portrait 832 x 1216
3:2 Landscape 1216 x 832
Custom... Opens a modal

Custom... takes up to 20 characters and accepts either a ratio like 21:9 or exact pixels like 1536x640. Your current custom value is shown as a pre-selected option labelled "Your custom ratio".

Warning

Custom aspect ratios are not checked when you save them. A typo is only caught on your next generation, which fails with a validation error instead.

Embeddings are filtered by your selected model, so an incompatible embedding is not listed. Both embedding dropdowns write to the same saved list, and picking in one dropdown only replaces the entries belonging to that dropdown.

The Prompts row has an Edit 📝 button that opens a modal with both prompt fields pre-filled. Clearing a field and saving removes that prompt. The modal does not touch the RAW toggle.

Warning

Turning RAW on throws away your saved default positive prompt as well as the model's own additions. RAW means your prompt text and nothing else. Your default negative prompt is still used.


Output#

Setting Values Default What it does
Format WebP or PNG WebP (inherited from the bot) File type for delivered images. WebP is smaller, PNG is lossless
Quality 1 to 100 90 WebP compression quality. The row is hidden when your format is PNG
Prompts Inline or Code Block Code Block How the prompt is printed on the result card
Previews On or Off On Live progress previews while a job renders. Off skips them entirely
Audio Format MP3, Opus, FLAC MP3 Audio file type for /compose
Audio Quality High, Medium, Low High Audio bitrate. The row is hidden when your format is FLAC, which is lossless

Format and both audio rows use Toggle / Cycle buttons. The Quality row uses an Edit 🎚️ button that opens a modal. Values that are not whole numbers between 1 and 100 are rejected.

Note

The Previews button is grey when previews are On, because that matches the bot default. It turns green when you have explicitly turned previews off.


Features#

Setting Values Default What it does
AI Disabled 🚫, Match Only 🎭, Enhance Only ✨, Both Enabled 🌟 Disabled Match enables automatic character detection in your prompt. Enhance makes AI prompt enhancement default-on
Match Sensitivity 0.0 to 1.0 0.85 Character-match cutoff. Only shown when AI is set to Match Only or Both
Re-Sample Denoise 0.02 to 0.99 0.30 Intended as the Detailer re-sample strength
History Per Page 9 or 18 9 Thumbnails per page in /history. 9 is one row, 18 is two stacked rows

The AI button cycles Disabled to Match Only to Enhance Only to Both and back.

Warning

Re-Sample Denoise is currently inert. The value saves and displays correctly, but nothing in the generation pipeline reads it, so changing it has no effect on your images.

Note

The Match Sensitivity modal and the on-page caption describe the direction in opposite terms. Treat 0.85 as the conservative baseline and adjust in small steps.


Describe#

Settings for the right-click describe context menu command.

Setting Values Default What it does
Backend Gemini 3 Flash, Gemini 3.1 Flash Lite, WD Tagger v3, JoyCaption Beta One Gemini 3.1 Flash Lite Which model writes the description
Output style Krea (natural language), Ideogram (text-aware), Booru (Danbooru tags), SD prompt (civitai mix), MidJourney prompt, Plain description Krea What the produced prompt looks like
Descriptions per run 1, 2, 3, 4 2 How many descriptions you get. Each one is a separate call, so higher costs more

The two Gemini backends are API-based and fast. WD Tagger v3 runs locally. JoyCaption Beta One is listed but not yet wired up.

The Gemini calls are billed to the maintainer's API key, so a backend can stop working the moment a quota runs out. WD Tagger v3 is the one that keeps going on local hardware.

Note

Booru output always uses WD Tagger v3 internally. If you pick Booru with a different backend, it silently falls back to the tagger.


Reset All#

Reset All 🔄 on the Home page opens a red confirmation card with Confirm Reset 🗑️ and Cancel ◀.

Confirming deletes your entire settings record: models, LoRAs, upscaler, prompts, aspect ratio, AI features, output and audio preferences, describe settings, sort methods, custom orders, and hidden items. Everything goes back to the system defaults. This is the only confirmation prompt anywhere in /settings.


Settings vs command parameters#

A parameter you type in a command always beats the matching saved setting, for that one run only. Your saved value is untouched.

Parameter Overrides
/imagine model: Imagine model
/imagine lora: and lora_strength: All four LoRA slots (see the warning below)
/imagine negative: Default negative prompt
/imagine aspect_ratio: Aspect ratio
/imagine nsfw: NSFW toggle
/imagine raw: RAW toggle
/imagine upscaler: Upscaler
/imagine enhance: The enhancement half of your AI setting
/upscale method: Upscale method
/edit quality: Edit model

Settings with no command parameter at all: default positive prompt, embeddings, every Output setting, match sensitivity, history page size, and every Describe setting. Those are only changeable in /settings.

lora: replaces your saved LoRAs

Passing /imagine lora: uses that single LoRA and blanks slots 2, 3, and 4 for that run, and it ignores the variant you saved for that LoRA. If you want your saved stack, leave lora: out.

RAW discards your default positive prompt

raw:True (or the saved RAW toggle) sends your prompt text and nothing else. Your saved default positive prompt is dropped along with the model's quality tags, embeddings, and LoRA triggers.

Note

Some settings are read at display time rather than generation time. The Reconstruct button in /history re-posts an old job using your current format, quality, and prompt-display settings, not the ones the job was originally made with.


There are no presets or profiles#

You have exactly one set of settings. There is no way to save, name, load, export, or share a bundle of settings, and there are no per-server or per-channel setting profiles. If you want a different configuration, you change your settings and change them back.

They are also not backed up anywhere you control. Everything here lives in one database file on the maintainer's machine, and a saved model or LoRA that gets removed from the bot quietly stops applying, so expect to reset things from time to time.

For one-off variations, use command parameters instead -- see Parameters.