Models#
There are 53 selectable image models loaded for /imagine at the moment. This page is
the master list, plus the rules that decide which one you get and what it costs you to
switch.
This list drifts
Models are added, swapped and removed whenever the maintainer feels like it, with no
announcement. A model documented here may simply not be in the picker any more, and a
model in the picker may not be here. The model: autocomplete in Discord is the only
list that is actually current.
Every model is available to every user. Picking one silently changes your steps, CFG, sampler, scheduler, resolution, quality tags, negative prompt, batch size and default upscaler all at once -- see Parameters.
How do I choose#
| You want | Go to | Best picks |
|---|---|---|
| Anime art from Danbooru tags | SDXL family | WAI (v16.0), mystery i32, JANKU (v6.9), Hassaku (v3.0) |
| Anime art from sentences | Anima, Z-Image | Anima Turbo, Z Anime Base, NetaYume AIO (v3.5) |
| Photoreal people and scenes | Krea 2, Flux | Krea, Flux Dev, Cyber Realistic Turbo |
| Text inside the image (signs, logos, labels) | Ideogram 4, Other models | Ideogram 4 (Quality), Qwen |
| Long prose prompts with lots of clauses | Krea 2, Ideogram 4, Flux | Krea, Ideogram 4, Klein 9B Quality, Mage Flow (Quality) |
| A result fast | /turbo |
/turbo is the fast path. Within /imagine, low step counts finish sooner: Klein 4B Fast (4 steps), Mage Flow Turbo (4), Z Anime 4-Step (4), ERNIE Turbo (8) |
| To change an image you already have | /edit |
a separate catalog of 14 edit models, not the 53 below |
Speed and the word \"turbo\"
A model having "turbo" in its name does not make it one of the fast models. The
name refers to how the checkpoint was built, not to how quickly it renders here.
What actually decides speed on /imagine is the step count, the resolution and how
busy the queue is.
The genuinely fast path is the /turbo command, which is a
separate pipeline with its own model list.
Tip
If you have no opinion, leave the model alone. The bot-wide default is Krea, and it is the family with the most tuning behind it.
The full catalog#
All 53, in the order the picker shows them. Steps and CFG are what actually runs when you do not override them. Base resolution is the size you get at the model's default aspect ratio.
| Model | Family | What it is for | Steps | CFG | Base res |
|---|---|---|---|---|---|
| Krea | Krea 2 | Default model. Photographic and illustrative, prose prompts, optional reference image | 12 | 1 | 1024x1024 |
| Z Image Turbo | Z-Image | Fast general-purpose Lumina2 build | 12 | 1 | 1024x1536 |
| Anima Turbo | Anima | Fast anime and illustration | 12 | 1 | 896x1152 |
| ERNIE Turbo | Other | Fast general-purpose Baidu model | 8 | 1 | 1024x1024 |
| Ideogram 4 | Ideogram 4 | Text rendering, photo-look scenes | 20 | 3 | 1216x1216 |
| mystery i32 | SDXL | Illustrious. Rigid default art style, very good character knowledge | 30 | 6 | 1152x1536 |
| FantasyAnime (v2.0) | SDXL | Illustrious anime | 30 | 5 | 1152x1536 |
| WAI (v16.0) | SDXL | Illustrious anime, general workhorse | 30 | 5 | 1152x1536 |
| One Obsession (v18) | SDXL | Illustrious anime | 20 | 7 | 1152x1536 |
| Cat Tower (v20 v-pred) | SDXL | NoobAI v-prediction | 30 | 6 | 896x1152 |
| Cat Carrier (v9.0) | SDXL | Illustrious anime | 30 | 6 | 1152x1536 |
| Echo Dream (v1.0) | SDXL | Illustrious, 8-step distilled -- fast SDXL | 12 | 1.5 | 1152x1536 |
| Cyber Realistic Pony (v16) | SDXL | Pony, realism leaning | 30 | 5 | 896x1152 |
| Milk (Walnut) | SDXL | Illustrious anime | 28 | 3 | 1152x1536 |
| Hassaku (v3.0) | SDXL | Illustrious anime | 20 | 7 | 1152x1536 |
| Dreambox (v4.0) | SDXL | Illustrious anime | 20 | 5 | 1152x1536 |
| Visionary (Surreal Chaos) | SDXL | Illustrious, surreal styling | 28 | 3.5 | 1152x1536 |
| Red Lily (Artsy) | SDXL | Illustrious, painterly | 30 | 5 | 1152x1536 |
| JANKU (v6.9) | SDXL | Illustrious, plug-and-play with default embeddings | 30 | 5 | 1152x1536 |
| Anima (aesthetic-v1.1) | Anima | Anime concepts, characters and styles, smoother output | 30 | 4 | 896x1152 |
| Anima Base (v1.0) | Anima | First full Anima release, highest step count | 50 | 5 | 896x1152 |
| Anima Cat Tower (v1.0) | Anima | Anima finetune | 30 | 5 | 896x1152 |
| Anima | Anima | Anima Preview 3 base | 30 | 5 | 896x1152 |
| AnimaYume (v0.5) | Anima | Anima Base finetune | 50 | 5 | 896x1152 |
| Cyber Realistic Turbo | Z-Image | Realism-leaning Lumina2 turbo, very large base size | 12 | 1 | 1280x1920 |
| ERNIE Image | Other | Baidu general-purpose, full quality pass | 20 | 4 | 1024x1024 |
| Klein 4B Fast | Flux | Flux 2 Klein 4B, fastest of the family | 4 | 1 | 1024x1024 |
| Klein 9B Fast | Flux | Flux 2 Klein 9B, fast | 4 | 1 | 1024x1024 |
| Klein 4B Quality | Flux | Flux 2 Klein 4B, full pass | 20 | 5 | 1024x1024 |
| Klein 9B Quality | Flux | Flux 2 Klein 9B, full pass | 20 | 5 | 1024x1024 |
| Flux Dev | Flux | Photorealistic, strong prompt following | 20 | 1.0 | 1024x1024 |
| Ideogram 4 (Fast) | Ideogram 4 | Distilled Ideogram, same text strengths, much faster | 20 | 1 | 1024x1024 |
| Ideogram 4 (Quality) | Ideogram 4 | Slowest, best text and detail | 48 | 3 | 2048x2048 |
| Ideogram 4 (Turbo) | Ideogram 4 | Fastest Ideogram, lower fidelity | 12 | 3 | 1024x1024 |
| ZiT-ANI | Z-Image | Lumina2 anime finetune | 12 | 1 | 1024x1536 |
| Krea (Quality) | Krea 2 | Krea sampled larger -- more VRAM, more detail | 12 | 1 | 1408x1408 latent |
| Krea (Exp) | Krea 2 | Raw Krea plus turbo LoRA. More composition variety per seed | 16 | 1.5 | 1024x1024 |
| Mage Flow (Quality) | Other | Microsoft Mage-Flow 4B, honours negative prompts | 20 | 5 | 1024x1024 |
| Mage Flow Turbo | Other | Microsoft Mage-Flow 4B, 4-step build | 4 | 1 | 1024x1024 |
| Mugen Aesthetic | Other | NoobAI weights on a Flux 2 VAE, flow-based sampling | 28 | 5 | 896x1152 |
| NetaYume AIO (v3.5) | Other | Neta-Lumina anime, instruction-style prompts | 40 | 4 | 864x1152 |
| Noob AI XL (vPred 1.0) | SDXL | NoobAI v-prediction base | 30 | 4 | 896x1152 |
| Nova Anime XL (v8) | SDXL | NoobAI anime | 25 | 5 | 896x1152 |
| Qwen | Other | Qwen Image INT4. Fast, strong text rendering | 20 | 2.5 | 1328x1328 |
| Quillworks (V2) | SDXL | Illustrious, simplified style | 35 | 5 | 1152x1536 |
| RedLilly (v1.0) | SDXL | Illustrious, cinematic styling | 30 | 5 | 1152x1536 |
| WAI Anima (v1.0) | Anima | WAI finetune of Anima. Ships its own hi-res-fix recipe | 30 | 5 | 896x1152 |
| waiNSFW (v13.0) | SDXL | Illustrious, older WAI branch | 35 | 5 | 1152x1536 |
| WAI-SHUFFLE (v-pred-04) | SDXL | NoobAI v-prediction | 30 | 5 | 896x1152 |
| Z Anime Base | Z-Image | Z-Image anime finetune, full quality pass | 35 | 4 | 1024x1536 |
| Z Anime 4-Step | Z-Image | Z Anime distilled, fastest | 4 | 1 | 1024x1536 |
| Z Anime 8-Step | Z-Image | Z Anime distilled, middle ground | 8 | 1 | 1024x1536 |
| Z Image Base | Z-Image | Z-Image full quality pass | 35 | 4 | 1024x1536 |
The description in the dropdown is not always right
Several models carry a one-line description that disagrees with what the sampler actually runs. Known cases: Krea and Krea (Quality) say "8 steps" but run 12; FantasyAnime (v2.0) says "28 steps, CFG 3.5" but runs 30 / 5; Hassaku (v3.0) says CFG 6 but runs 7; JANKU (v6.9) and waiNSFW (v13.0) say CFG 6 but run 5; Noob AI XL (vPred 1.0) says 32 steps but runs 30. The numbers in the table above are the ones that run.
Note
Long descriptions are cut at 100 characters in the dropdown, so Krea, Ideogram 4 (Fast), Krea (Exp) and Mage Flow (Quality) appear truncated mid-sentence.
The two-level picker#
The 53 models do not appear as 53 flat options. They collapse into 35 groups:
- 10 multi-version families that expand when you click them: Krea 2, Z Image, Anima, ERNIE, Ideogram, WAI, Red Lily, Klein, Mage Flow, Z Anime. These 10 hold 28 of the 53 models between them.
- 25 direct picks that select a model in one click.
A family option is labelled with the family name and a subtitle reading
"N versions - latest: <name>". Choosing it does not start a render -- it opens a labelled
button list underneath, headed <Family> versions - newest first, with one Select
button per version. The version you are already on shows as a green, disabled Selected
button.
Discord caps a dropdown at 25 options, so 35 groups need 2 pages. A < / N/M / >
row appears under the list to move between them. Anything the panel adds above the models
(for example the detailer's Inherit option) eats into that 25-option budget and can push
the split around.
Inside a family, versions sort newest-first: by priority, then version number descending, with named alternates like Base, Turbo and Quality floated to the top of a tie.
Three families look like single models
Cat Tower, Dreambox and Flux Dev are declared as families but currently have only one enabled version each, so they render as plain direct picks with no drill-in.
The model: option on the /imagine slash command works differently: it is a flat
autocomplete. With no text typed it offers the first 25 models one per row. As you type it
matches on family name, nickname or filename, and surfaces up to 4 sibling versions per
family so alternates stay visible inline.
You can reorder or hide models for yourself in /settings --
sort by priority (the default), alphabetically, or a custom order, and hide anything you
never use. That only affects your own picker.
Access and gating#
No image model is VIP-gated or admin-gated. All 53 are available to every user who can use the bot at all. There is no NSFW-locked model, no premium tier, no per-server model list. A model only disappears if the operator disables it or you hide it yourself.
The one exception is editing
The /edit command draws on a separate catalog of 14 edit
models, and 8 of those are VIP-only: every nano* variant (Nano Banana, Nano
Banana 2, Nano Banana 2 Lite, Nano Banana Pro) and every gpt* variant. They run on
the maintainer's own paid API keys, which is the only reason they are gated. The other
six (klein, qwen, krea2, mage, restyle, z_image) are open to everyone. VIP
status is checked live, so a promotion takes effect on your next /edit.
NSFW is a setting, not a model list#
nsfw is a per-user toggle in /settings. It does not unlock or
block any model. What it changes is the guard text merged into your prompt:
Your nsfw setting |
What happens |
|---|---|
| On | Nothing is added. Only the model's own baked-in positive and negative text applies. |
| Off | The model's safe-content strings are merged in -- its SFW negative additions go on the negative prompt and its SFW positive additions go on the positive prompt. |
The exact strings are per-model. Typical SFW negative additions are nsfw, explicit, nude,
naked, sex, penetration, sometimes weighted like (nsfw:1.2); typical SFW positive
additions are safe or (safe:1.2). waiNSFW (v13.0) uses the strongest weighting in the
catalog at (nsfw:1.4), (explicit:1.4).
Roughly half the catalog carries guard strings at all. The pure DiT models -- Krea 2, Ideogram 4, Klein, Flux Dev, Qwen, Mage Flow, ERNIE, Z Image and Z Anime -- ship none, so turning the setting off changes nothing on those.
Warning
raw: true on /imagine discards every merged string --
quality tags, SFW guards, LoRA triggers and enhancement. Only your literal text is
encoded, whatever your NSFW setting says.
Licensing and credits#
Every model, LoRA, upscaler and encoder a job touched is recorded on the job itself at the moment it is created, so old results keep correct credits even after the catalog changes. Re-rolls, upscales and edits accumulate the chain, so a heavily worked image lists everything that ever touched it.
Credits are not stamped on the image. They appear in the 🔍 Details view of a result, on
a small-text models line where each entry links to its source and its license. See
Details panel.
25 licenses are defined, grouped into six categories and ranked by strictness:
| Category | Priority (1 = strictest) | Examples |
|---|---|---|
| API Terms | 1 | Google Terms of Service, OpenAI Terms of Use |
| Proprietary | 1 | Proprietary |
| Non-Commercial | 2 | CC BY-NC 4.0, CC BY-NC-SA 4.0, FLUX.1 [dev], CreativeML Open RAIL-M (NC) |
| Copyleft / Restricted | 3 | GPL 3.0, AGPL 3.0, CC BY-SA 4.0, Krea 2 Community License, Kohaku 1.0, Gemma Terms |
| Attribution | 4 | Apache 2.0, MIT, CC BY 4.0, FAIPL-1.0-SD, CreativeML Open RAIL-M |
| Permissive | 5 | CC0, Unlicense, WTFPL |
When a job used several models, the most restrictive license wins the summary slot, and
the full list is de-duplicated and sorted strictest-first. Anything with no license declared
resolves to Unknown.
Most SDXL models do not declare a license explicitly -- it is inferred from the subtype:
illustrious, noob and pony all resolve to FAIPL-1.0-SD, and lumina2 resolves to
Apache 2.0. Across the visible catalog that works out to 27 models under FAIPL-1.0-SD,
13 under Apache 2.0, 3 under the Krea 2 Community License, 2 under MIT and the rest
non-commercial.
Two license names are not defined
The 4 Ideogram entries and Flux Dev point at license identifiers that have no entry in the license table, so they render without a proper name or link in the Details view. Treat both as non-commercial and check the model's own source page.
Family pages#
| Page | Covers |
|---|---|
| SDXL family | 20 models -- Illustrious, NoobAI, Pony, plain SDXL, v-pred |
| Z-Image | 7 models -- Z Image, Z Anime, Cyber Realistic Turbo, ZiT-ANI |
| Krea 2 | 3 models -- the default family |
| Flux | 5 models -- Flux 2 Klein 4B/9B and Flux Dev |
| Ideogram 4 | 4 models -- the text-rendering family |
| Anima | 7 models -- Anima and its finetunes |
| Other models | 7 models -- ERNIE, Mage Flow, Mugen, NetaYume, Qwen |
| Upscaler catalog | The upscale models, not generators |