SDXL Family#
Twenty of Eimi's 53 models are SDXL checkpoints. This is the classic Danbooru-tag anime path
-- the models that want 1girl, long hair, blue eyes, school uniform rather than a sentence.
If you are new and want anime art, start here. If you want a photo or text in the image, go to Krea 2 or Ideogram 4 instead.
This catalog moves
Models get added, swapped and disabled whenever the maintainer feels like it. The counts and defaults below were true when this page was written; the model picker is the authority.
What makes SDXL different#
Four things this family does that the newer DiT families do not:
| Behaviour | What it means for you |
|---|---|
| Automatic quality tags | Most models silently prepend or append their own positive tags (usually masterpiece, best quality) and add their own negative list. You never type them. |
| Real negative prompts | negative: on /imagine actually does something here. On Krea 2, Ideogram 4, Klein, Flux Dev and ERNIE it does not. |
| Embeddings | Several models load textual-inversion embeddings by default -- lazypos, lazyneg, lazynsfw, deep_negative_pony, smooth_negative. These are gated to Illustrious, plain SDXL and Pony types and are never loaded on other families. |
| TIPO enhancement | All 20 run their AI enhance through TIPO, a small tag-expansion model that runs inside ComfyUI rather than calling an external LLM. Enhancement is opt-in here: nothing happens unless you pass enhance:. |
The family also has the deepest LoRA library: 28 of the 91 LoRAs are SDXL-compatible, against 44 for Z-Image and 5 for Krea 2. Up to 4 LoRAs stack at once and they affect both the model and the text encoder. See LoRAs.
Prompt these models with tags
Comma-separated Danbooru tags, most important first. Weights use (tag:1.2) syntax.
Aim for 20 to 35 tags -- SDXL's text encoder only reads about 75 tokens, so a long
paragraph loses its tail. Full guidance in Prompting.
Subtypes#
Every live SDXL model belongs to one of three subtypes. The subtype decides your default resolution and how character matching writes character names into your prompt.
| Subtype | Live models | Base resolution | Character name format |
|---|---|---|---|
| Illustrious | 15 | 1152x1536 (~1.77 MP) | hatsune_miku \(vocaloid\) |
| NoobAI | 4 | 896x1152 (~1.03 MP) | hatsune_miku \(vocaloid\) |
| Pony | 1 | 896x1152 | hatsune_miku \(vocaloid\) |
| Plain SDXL | none currently enabled | 896x1152 | -- |
| v-pred | a modifier, not a bucket | inherited from the base subtype | -- |
Illustrious is the newest of the three lineages and renders at roughly 70 percent more pixels by default, so it is slower per image but sharper out of the box.
NoobAI and Pony run at the older ~1 MP SDXL size. Pony models also expect the
score_9, score_8_up, score_7_up prompt convention, which Cyber Realistic Pony adds for you.
v-pred (v-prediction) is a different noise-prediction objective, not a different style family. In practice a v-pred checkpoint holds true blacks and high contrast far better than an epsilon one, but it is fussier: the sampler and scheduler it ships with are part of the recipe, and swapping them (or letting a LoRA swap them for you) can wash the image out or blow the highlights. Two models declare it: Cat Tower (v20 v-pred) and Noob AI XL (vPred 1.0). The other 18 are epsilon models and shrug off sampler changes.
Note
WAI-SHUFFLE (v-pred-04) is named for a v-pred training run but is registered as a plain NoobAI model, so it gets no v-pred handling and follows the NoobAI resolution bucket like the rest.
Mugen Aesthetic is a near relative: SDXL weights running on a Flux 2 VAE with flow-based sampling. It uses the same 896x1152 NoobAI bucket but is built by a different pipeline, so it lives on the Other models page.
All 20 at a glance#
Steps and CFG below are what runs when you override nothing. Full spec tables are in the per-model sections underneath.
| Model | Subtype | Steps | CFG | Base res | Best for |
|---|---|---|---|---|---|
| mystery i32 | Illustrious | 30 | 6 | 1152x1536 | Character accuracy with a fixed house style |
| FantasyAnime (v2.0) | Illustrious | 30 | 5 | 1152x1536 | Anime screencap look, detailed backgrounds |
| WAI (v16.0) | Illustrious | 30 | 5 | 1152x1536 | The general workhorse, light baked prompt |
| One Obsession (v18) | Illustrious | 20 | 7 | 1152x1536 | Fast, obedient, depth-of-field portraits |
| Cat Carrier (v9.0) | Illustrious | 30 | 6 | 1152x1536 | Clean high-detail anime with artifacts suppressed |
| Echo Dream (v1.0) | Illustrious | 12 | 1.5 | 1152x1536 | Speed -- the only distilled SDXL here |
| Milk (Walnut) | Illustrious | 28 | 3 | 1152x1536 | Soft low-CFG rendering, loose interpretation |
| Hassaku (v3.0) | Illustrious | 20 | 7 | 1152x1536 | Fast and literal, minimal baked prompt |
| Dreambox (v4.0) | Illustrious | 20 | 5 | 1152x1536 | Fast SDE sampling with extra texture |
| Visionary (Surreal Chaos) | Illustrious | 28 | 3.5 | 1152x1536 | Surreal, dreamlike, low-CFG compositions |
| Red Lily (Artsy) | Illustrious | 30 | 5 | 1152x1536 | Painterly art focus with heavy anatomy cleanup |
| JANKU (v6.9) | Illustrious | 30 | 5 | 1152x1536 | Plug and play -- embeddings do the work |
| Quillworks (V2) | Illustrious | 35 | 5 | 1152x1536 | Painted illustration rather than flat anime |
| RedLilly (v1.0) | Illustrious | 30 | 5 | 1152x1536 | Cinematic lighting and blurred backgrounds, unasked |
| waiNSFW (v13.0) | Illustrious | 35 | 5 | 1152x1536 | The older WAI branch, highest step ceiling |
| Cat Tower (v20 v-pred) | NoobAI, v-pred | 30 | 6 | 896x1152 | Dark scenes and true blacks |
| Noob AI XL (vPred 1.0) | NoobAI, v-pred | 30 | 4 | 896x1152 | The reference NoobAI v-pred base |
| Nova Anime XL (v8) | NoobAI | 25 | 5 | 896x1152 | Scenery and landscapes, fastest NoobAI |
| WAI-SHUFFLE (v-pred-04) | NoobAI | 30 | 5 | 896x1152 | Newest-era Danbooru aesthetic, mixed styles |
| Cyber Realistic Pony (v16) | Pony | 30 | 5 | 896x1152 | Semi-realistic output, widest CFG range |
Steps and CFG are clamped, not rejected
Every model declares a recommended range. Pass something outside it and you get an image at the nearest legal value rather than an error. See Parameters.
Five dropdown descriptions are wrong
The subtitle text in the model picker disagrees with what actually runs for these:
| Model | Dropdown says | Actually runs |
|---|---|---|
| FantasyAnime (v2.0) | 28 steps, CFG 3.5 | 30 steps, CFG 5 |
| Hassaku (v3.0) | 20 steps, CFG 6 | 20 steps, CFG 7 |
| JANKU (v6.9) | 30 steps, CFG 6 | 30 steps, CFG 5 |
| waiNSFW (v13.0) | 35 steps, CFG 6 | 35 steps, CFG 5 |
| Noob AI XL (vPred 1.0) | 32 steps, CFG 4 | 30 steps, CFG 4 |
Trust the spec tables below, not the dropdown.
Two things are the same on all 20, so they are not repeated in prose: the default batch is 2, and the default upscaler is AnimeSharp (4x) because no SDXL model overrides the global one. Both are changeable per job.
Illustrious models#
Fifteen of the twenty. All render at 1152x1536 by default and all take Danbooru tags.
mystery i32#
An unlabelled merge with a fixed house style and unusually good character recall. Pick it when you want a named character to come out right and you do not care about steering the art style much.
| Architecture | SDXL |
| Subtype | Illustrious |
| Prompting | Danbooru tags, 20-35 of them |
| Negatives | Supported. Two negative embeddings are always loaded |
| Steps | 30 (25-50) |
| CFG | 6 (2-9) |
| Sampler | euler_ancestral |
| Scheduler | normal |
| Base resolution | 1152x1536 |
| Default batch | 2 |
| Default upscaler | AnimeSharp (4x) |
Good at: named-character accuracy, a consistent look with no style tags at all, and wide tuning latitude -- 25 to 50 steps and CFG 2 to 9 is one of the two loosest ranges in the family.
Weak at: style variety. Its own description calls the default art style "rigid", and artist tags fight it. It also ships no positive quality tags, so a thin prompt renders flat where WAI or Cat Carrier would have propped it up. The two always-on negative embeddings sand down deliberately grungy, sketchy or lo-fi looks. At 30 steps and 1.77 MP it is not a fast model.
Notes: the always-loaded embeddings are deep_negative_pony (strength 1.0) and
smooth_negative (strength 0.8), both on the negative side.
FantasyAnime (v2.0)#
A screencap-styled Illustrious merge that carries the heaviest built-in prompt in the family on both sides. Good when you want the anime-episode look without describing it.
| Architecture | SDXL |
| Subtype | Illustrious |
| Prompting | Danbooru tags, 20-35 of them |
| Negatives | Supported. Long built-in list plus lazyneg and lazyhand |
| Steps | 30 (15-30) |
| CFG | 5 (5-7) |
| Sampler | euler |
| Scheduler | normal |
| Base resolution | 1152x1536 |
| Default batch | 2 |
| Default upscaler | AnimeSharp (4x) |
Good at: anime screencap and cel-shaded output with no style prompting, hands (it loads
lazyhand and weights (5fingers:1.6) into the negative), and busy detailed backgrounds
(it actively pushes (simple background:1.6) into the negative).
Weak at: flat, minimal or transparent backgrounds and negative space -- you are fighting
its own negative to get them. Sketch and lineart styles are suppressed the same way, as are
sweat, saliva and wet-look tags. euler is deterministic, so seeds vary less than on the
euler_ancestral models. The 15-30 step range gives you no long-schedule option.
Notes: its quality tags are appended after your prompt rather than prepended, so your own leading tags keep their weight. The picker subtitle for this one is wrong -- see the warning above.
WAI (v16.0)#
The newest WAI branch and the closest thing this family has to a default pick. Light built-in prompt, forgiving range, no strong opinion about style.
| Architecture | SDXL |
| Subtype | Illustrious |
| Prompting | Danbooru tags, 20-35 of them |
| Negatives | Supported. Short built-in list |
| Steps | 30 (15-30) |
| CFG | 5 (5-7) |
| Sampler | euler_ancestral |
| Scheduler | normal |
| Base resolution | 1152x1536 |
| Default batch | 2 |
| Default upscaler | AnimeSharp (4x) |
Good at: general-purpose anime across a wide subject range, letting your tags dominate (its built-in prompt is four words on each side), and running fast when you drop it to 15 steps.
Weak at: having a look of its own -- a plain prompt gives you a generic result, which is the price of the light baked prompt. The short negative list means anatomy errors are not caught for you the way FantasyAnime or Red Lily catch them. No photorealism.
Notes: shares a picker group with waiNSFW (v13.0). This is the newer of the two.
One Obsession (v18)#
A fast, high-CFG aesthetic model that bakes depth of field into every render. Twenty steps makes it one of the quicker full-quality Illustrious options.
| Architecture | SDXL |
| Subtype | Illustrious |
| Prompting | Danbooru tags, 20-35 of them |
| Negatives | Supported. Built-in anatomy and watermark list |
| Steps | 20 (15-25) |
| CFG | 7 (5-8) |
| Sampler | euler |
| Scheduler | normal |
| Base resolution | 1152x1536 |
| Default batch | 2 |
| Default upscaler | AnimeSharp (4x) |
Good at: speed at full resolution, literal tag obedience (CFG 7 default, up to 8), and portrait-style shots with shallow depth of field, which it adds for you.
Weak at: flat or simple backgrounds and backlit or silhouette shots -- simple
background, transparent and face backlighting are all in its built-in negative. Text and
logos are suppressed too. CFG 7 by default means colour burn and contrast crush on prompts
that were already saturated, and the 15-25 step ceiling caps how far you can refine.
Cat Carrier (v9.0)#
A general anime merge tuned for clean output. Tagged in the picker as a base model and a checkpoint merge.
| Architecture | SDXL |
| Subtype | Illustrious |
| Prompting | Danbooru tags, 20-35 of them |
| Negatives | Supported. Built-in artifact list |
| Steps | 30 (20-30) |
| CFG | 6 (5-7) |
| Sampler | euler_ancestral |
| Scheduler | normal |
| Base resolution | 1152x1536 |
| Default batch | 2 |
| Default upscaler | AnimeSharp (4x) |
Good at: clean, high-detail anime illustration, and suppressing junk -- jpeg artifacts, scan artifacts and watermarks are all in the built-in negative. It handles a wide subject range without steering the style.
Weak at: sparkle, bokeh and particle-heavy fantasy effects, because light particles
sits in its negative and you are fighting it. Sketch styles are pushed away too. As a merge
it has no distinctive look, so it rarely beats a specialist model at that model's job.
Echo Dream (v1.0)#
The only distilled model in the family. Twelve steps at CFG 1.5, so it renders in a fraction of the time of everything else here.
| Architecture | SDXL |
| Subtype | Illustrious |
| Prompting | Danbooru tags, 20-35 of them |
| Negatives | Accepted, but weak at CFG 1.5 |
| Steps | 12 (8-15) |
| CFG | 1.5 (1.2-1.8) |
| Sampler | euler_ancestral |
| Scheduler | simple |
| Base resolution | 1152x1536 |
| Default batch | 2 |
| Default upscaler | AnimeSharp (4x) |
Good at: speed, exploring a lot of seeds cheaply, and acting as the base pass for a hi-res fix or detailer run where the first pass does not need to be perfect.
Weak at: negative prompts -- at CFG 1.5 there is very little guidance signal for them to push against, so both your negative and its built-in one barely register. Complex multi-part compositions come out muddled compared to a 30-step model, and fine texture detail is visibly softer.
Do not raise the steps
Distilled models are trained for their short schedule. Anything above 15 steps or CFG 1.8 is clamped back to the boundary, and even inside the range, pushing CFG toward 1.8 tends to produce contrast artifacts rather than more detail.
Milk (Walnut)#
The Walnut flavour of the Plant Milk suite. A low-CFG model that interprets rather than transcribes your prompt.
| Architecture | SDXL |
| Subtype | Illustrious |
| Prompting | Danbooru tags, 20-35 of them |
| Negatives | Supported. Short built-in list |
| Steps | 28 (25-35) |
| CFG | 3 (2-6) |
| Sampler | euler_ancestral |
| Scheduler | sgm_uniform |
| Base resolution | 1152x1536 |
| Default batch | 2 |
| Default upscaler | AnimeSharp (4x) |
Good at: soft, low-contrast rendering that does not look over-guided, and a genuinely wide CFG range (2-6) so you can dial obedience up when a prompt needs it.
Weak at: literal adherence at its default CFG 3 -- long tag lists get partially ignored and small details drift. Sharp graphic linework is not its strength. Only the Walnut variant of the suite is loaded, so if the flavour is wrong there is no sibling to switch to.
Hassaku (v3.0)#
A fast, high-CFG Illustrious model with almost no built-in prompt. The one to pick when you want your own tags to be the whole instruction.
| Architecture | SDXL |
| Subtype | Illustrious |
| Prompting | Danbooru tags, 20-35 of them |
| Negatives | Supported. Three-item built-in list |
| Steps | 20 (20-30) |
| CFG | 7 (5-7) |
| Sampler | euler |
| Scheduler | normal |
| Base resolution | 1152x1536 |
| Default batch | 2 |
| Default upscaler | AnimeSharp (4x) |
Good at: speed at 20 steps, strong tag obedience at CFG 7, and staying out of your way --
its built-in negative is just worst quality, bad quality, low quality.
Weak at: soft or subtle styles, which CFG 7 tends to harden. The near-empty negative
means nothing is protecting you from bad hands or anatomy unless you write it yourself.
euler is deterministic, so a batch varies less than on the ancestral models, and CFG is
capped at 7 so there is no headroom above the default.
Notes: the picker subtitle says CFG 6. It runs at 7.
Dreambox (v4.0)#
A 20-step model on an SDE sampler, which adds noise back in at each step and gives slightly grainier, more varied output than the euler models.
| Architecture | SDXL |
| Subtype | Illustrious |
| Prompting | Danbooru tags, 20-35 of them |
| Negatives | Supported. Three-item built-in list |
| Steps | 20 (20-25) |
| CFG | 5 (4-6) |
| Sampler | dpmpp_2m_sde |
| Scheduler | sgm_uniform |
| Base resolution | 1152x1536 |
| Default batch | 2 |
| Default upscaler | AnimeSharp (4x) |
Good at: fast renders with extra surface texture, and a balanced CFG 5 default that suits both loose and tight prompts.
Weak at: tuning room. Its 20-25 step and 4-6 CFG ranges are the narrowest pair on any
Illustrious model here, so there is nowhere to go when the default is not working. The SDE
sampler injects fresh noise, which means re-running the same seed drifts more than it would
on euler. The three-word negative offers no anatomy protection.
Visionary (Surreal Chaos)#
The dedicated weird one. Low CFG, wide step range, and a very narrow CFG band that keeps it from ever becoming a literal model.
| Architecture | SDXL |
| Subtype | Illustrious |
| Prompting | Danbooru tags, 20-35 of them |
| Negatives | Supported. Built-in artifact and anatomy list |
| Steps | 28 (20-35) |
| CFG | 3.5 (3-4) |
| Sampler | euler_ancestral |
| Scheduler | simple |
| Base resolution | 1152x1536 |
| Default batch | 2 |
| Default upscaler | AnimeSharp (4x) |
Good at: surreal, dreamlike and abstract compositions, unexpected colour and shape choices, and long schedules -- 35 steps is available if you want to let the chaos settle.
Weak at: doing what you asked. Its CFG band tops out at 4, the tightest ceiling in the family, so tag adherence is loose by design, character likenesses drift, and clean commercial-looking illustration is not on the menu.
CFG above 4 is clamped
Passing cfg: 7 to this model gets you CFG 4. If you want obedience, use a different
model rather than fighting this one.
Red Lily (Artsy)#
The painterly branch of the Red Lily pair. Sampler settings are identical to RedLilly (v1.0); the difference is entirely in the built-in prompt, which here is short on the positive side and long on the negative.
| Architecture | SDXL |
| Subtype | Illustrious |
| Prompting | Danbooru tags, 20-35 of them |
| Negatives | Supported. Long built-in anatomy list |
| Steps | 30 (28-35) |
| CFG | 5 (4-5) |
| Sampler | euler_ancestral |
| Scheduler | normal |
| Base resolution | 1152x1536 |
| Default batch | 2 |
| Default upscaler | AnimeSharp (4x) |
Good at: art-focused illustration where you want resolution and aesthetics tags without a prescribed lighting style, and anatomy cleanup -- its built-in negative covers extra limbs, cloned faces, distorted proportions and duplicated legs.
Weak at: volumetric lighting and hand close-ups, which sit in its own negative list (see below). Muscular female body types are pushed away as well. The CFG 4-5 band is narrow, so there is no way to force stronger adherence on a difficult prompt.
Two things you might want are in its negative list
volumetric lighting and detailed hands are both baked into this model's negative
prompt. Asking for either in your positive prompt puts you in a tug of war. Use
RedLilly (v1.0) if you want volumetric lighting handled for you instead.
JANKU (v6.9)#
A merge that leans entirely on embeddings for its quality handling. Its own description calls it very plug and play, and that is accurate: your prompt is nearly the only text in the box.
| Architecture | SDXL |
| Subtype | Illustrious |
| Prompting | Danbooru tags, 20-35 of them |
| Negatives | Supported. Built-in negative is the lazyneg embedding only |
| Steps | 30 (25-30) |
| CFG | 5 (3-6) |
| Sampler | euler_ancestral |
| Scheduler | simple |
| Base resolution | 1152x1536 |
| Default batch | 2 |
| Default upscaler | AnimeSharp (4x) |
Good at: working straight out of the box with a short prompt, keeping your token budget
free (the lazypos and lazyneg embeddings each cost one token instead of a dozen quality
words), and a fairly wide CFG 3-6 band.
Weak at: letting you opt out. The embeddings apply on every job whether they suit the image or not, and they carry their own aesthetic bias. The 25-30 step band is tight, and because no literal quality words are added, a lazy prompt has less to lean on than on WAI or Cat Carrier.
Notes: despite being registered as Illustrious, its name references a NoobAI and Rouwei training base. Practically that only matters in that it uses the Illustrious resolution bucket. The picker subtitle says CFG 6; it runs at 5.
Quillworks (V2)#
An experimental build that deliberately pulls away from flat anime toward a painted illustration look. Widest tuning range in the family alongside mystery i32.
| Architecture | SDXL |
| Subtype | Illustrious |
| Prompting | Danbooru tags, 20-35 of them |
| Negatives | Supported. Built-in list, no built-in positive tags |
| Steps | 35 (25-50) |
| CFG | 5 (2-9) |
| Sampler | euler_ancestral |
| Scheduler | simple |
| Base resolution | 1152x1536 |
| Default batch | 2 |
| Default upscaler | AnimeSharp (4x) |
Good at: painted and illustrated output rather than cel shading, and experimentation -- 25 to 50 steps and CFG 2 to 9 gives you more room than almost anything else here. It also has the joint-highest default step count (35).
Weak at: flat anime, monochrome manga pages, and multi-panel or multi-view layouts, all of which are in its built-in negative. It adds no positive quality tags, so weak prompts get no help. It is an experimental checkpoint, so its behaviour is the most likely on this page to change or disappear.
This model's negative prompt contains the word anime
Along with shiny, plastic, monochrome, multiple views and child. That is
intentional -- it is how the painted look is enforced -- but it means asking for a clean
anime style here works against the model. Use WAI (v16.0) or
Cat Carrier (v9.0) for that.
RedLilly (v1.0)#
The other half of the Red Lily pair, and near-identical to Red Lily (Artsy) on paper: same steps, CFG, sampler, scheduler and ranges. The difference is the built-in prompt, which here is a long cinematic block.
| Architecture | SDXL |
| Subtype | Illustrious |
| Prompting | Danbooru tags, 20-35 of them |
| Negatives | Supported. Long built-in anatomy and body-type list |
| Steps | 30 (28-35) |
| CFG | 5 (4-5) |
| Sampler | euler_ancestral |
| Scheduler | normal |
| Base resolution | 1152x1536 |
| Default batch | 2 |
| Default upscaler | AnimeSharp (4x) |
Good at: cinematic shots with no lighting prompt at all -- dramatic lighting, volumetric
lighting, atmosphere, dynamic angles and depth of field are all added for you, along with an
(impressionism:1.5) weight.
Weak at: flat, graphic or sharply focused compositions. Its built-in positive forces a blurred background and an impressionist bias at high weight, so a crisp full-focus image is hard to get. Body-type variety is limited, since its negative pushes away heavier builds. Neutral or clinical lighting is effectively unavailable.
Notes: if the cinematic block is fighting you, switch to Red Lily (Artsy) -- everything else about the two is the same.
waiNSFW (v13.0)#
The older WAI branch, kept alongside WAI (v16.0) in the same picker group. Same built-in prompt as v16, and the highest step ceiling in the family.
| Architecture | SDXL |
| Subtype | Illustrious |
| Prompting | Danbooru tags, 20-35 of them |
| Negatives | Supported. Short built-in list |
| Steps | 35 (25-40) |
| CFG | 5 (5-7) |
| Sampler | euler_ancestral |
| Scheduler | normal |
| Base resolution | 1152x1536 |
| Default batch | 2 |
| Default upscaler | AnimeSharp (4x) |
Good at: long refinement schedules -- 35 steps by default and up to 40, more than any other Illustrious model here -- and giving you a second WAI aesthetic when v16's rendering of a prompt is not what you wanted.
Weak at: speed. Thirty-five steps at 1.77 MP is the slowest default here, tied with Quillworks (V2). It is also the older branch, so v16 generally has the more current knowledge. Like v16 its short negative gives no anatomy protection.
Strongest safe-content weighting in the catalog
With your NSFW setting off, this model applies (nsfw:1.4), (explicit:1.4) to the
negative -- higher than any other model. That is enough to visibly change poses, framing
and clothing even on prompts that were never near the line. If a SFW render comes out
oddly stiff, this is why. The picker subtitle also says CFG 6; it runs at 5.
NoobAI models#
Four models on the older 896x1152 bucket. Two are v-pred and two are epsilon. All take the same Danbooru tags as the Illustrious models but were trained on a different data mix, so tag coverage and era feel differ.
Cat Tower (v20 v-pred)#
A v-prediction NoobAI merge. Reach for it when a scene is meant to be dark and you keep getting grey.
| Architecture | SDXL |
| Subtype | NoobAI, v-pred |
| Prompting | Danbooru tags, 20-35 of them |
| Negatives | Supported. Built-in artifact list |
| Steps | 30 (25-30) |
| CFG | 6 (5-7) |
| Sampler | euler |
| Scheduler | normal |
| Base resolution | 896x1152 |
| Default batch | 2 |
| Default upscaler | AnimeSharp (4x) |
Good at: true blacks, night scenes and high-contrast lighting, which is what v-pred training buys you. Suppresses scan and jpeg artifacts and stray light particles via its built-in negative.
Weak at: fine detail compared to the Illustrious models, simply because it renders about 40 percent fewer pixels by default. The 25-30 step band is tight. Sketch styles are in the negative. And it is the fussier prediction type, so it is less tolerant of changes you make to the sampling recipe.
Leave the sampler alone on v-pred models
The euler / normal pairing is part of this model's recipe. Overriding the sampler or
scheduler -- or picking a LoRA that overrides them for you, which happens silently --
is far more likely to produce a washed-out or blown-out image here than on an epsilon
model. Check the 🔍 Details view if a render looks wrong.
Noob AI XL (vPred 1.0)#
The reference NoobAI v-pred checkpoint rather than a merge of it. The most predictable member of the NoobAI group and the base a lot of the others descend from.
| Architecture | SDXL |
| Subtype | NoobAI, v-pred |
| Prompting | Danbooru tags, 20-35 of them |
| Negatives | Supported. Built-in quality and anthro list |
| Steps | 30 (28-35) |
| CFG | 4 (4-5) |
| Sampler | euler |
| Scheduler | normal |
| Base resolution | 896x1152 |
| Default batch | 2 |
| Default upscaler | AnimeSharp (4x) |
Good at: broad Danbooru tag and artist coverage, dark and high-contrast scenes, and colour stability -- the tight CFG 4-5 band keeps it from burning the way a CFG 7 model can.
Weak at: anthro, furry and animal-character subjects. mammal, anthro, furry, ambiguous
form, semi-anthro are all in its built-in negative, and feral is added when your NSFW
setting is off, so this is the wrong model for that work. The CFG band is the narrowest of
the NoobAI four, and the same v-pred sampler caution applies as for Cat Tower.
Notes: the picker subtitle says 32 steps. It runs at 30.
Nova Anime XL (v8)#
An epsilon NoobAI model with a long built-in positive aimed at scenery. The fastest NoobAI option at 25 steps.
| Architecture | SDXL |
| Subtype | NoobAI |
| Prompting | Danbooru tags, 20-35 of them |
| Negatives | Supported. Long built-in list |
| Steps | 25 (20-30) |
| CFG | 5 (4-6) |
| Sampler | euler_ancestral |
| Scheduler | normal |
| Base resolution | 896x1152 |
| Default batch | 2 |
| Default upscaler | AnimeSharp (4x) |
Good at: landscapes and environment shots -- scenery, depth of field and volumetric
lighting are all in its built-in positive. Fastest of the NoobAI four, and being epsilon it
takes sampler and scheduler changes without complaint.
Weak at: anything that is not anime illustration. cartoon, graphic, painting,
crayon, graphite and abstract are all in its built-in negative, so traditional-media
looks are actively suppressed. simple background is negative too. Era tags are a trap: it
negates modern, recent, old and oldest at once, so period styling is unreliable.
WAI-SHUFFLE (v-pred-04)#
A style-shuffling NoobAI merge aimed at the newest-era Danbooru aesthetic. Named for a v-pred run but registered as a plain epsilon NoobAI model.
| Architecture | SDXL |
| Subtype | NoobAI |
| Prompting | Danbooru tags, 20-35 of them |
| Negatives | Supported. Built-in list |
| Steps | 30 (28-35) |
| CFG | 5 (4-5) |
| Sampler | euler_ancestral |
| Scheduler | normal |
| Base resolution | 896x1152 |
| Default batch | 2 |
| Default upscaler | AnimeSharp (4x) |
Good at: current-era anime aesthetics (its built-in positive includes newest,
absurdres, highres, very awa), varied style output across seeds, and safe sampler
overrides since it is an epsilon model despite the name.
Weak at: retro and early-era anime looks -- old and early are in its negative. Named
artist styles are working uphill too, since artist name, copyright name and signature
are all negated. The CFG 4-5 band is narrow, and it shares the NoobAI resolution ceiling.
Notes: the "v-pred" in the name is the upstream training run, not a flag the bot acts on. There is no v-pred handling and no special sampler caution here.
Pony models#
One model, on the 896x1152 bucket, using the score_* prompt convention instead of plain
quality tags.
Cyber Realistic Pony (v16)#
The only Pony checkpoint loaded, and the only realism-leaning model in the whole SDXL family.
It is also the only one on dpmpp_2m / karras.
| Architecture | SDXL |
| Subtype | Pony |
| Prompting | Danbooru tags with the score_* chain, which is added for you |
| Negatives | Supported. Built-in score and anatomy list |
| Steps | 30 (30-40) |
| CFG | 5 (2-10) |
| Sampler | dpmpp_2m |
| Scheduler | karras |
| Base resolution | 896x1152 |
| Default batch | 2 |
| Default upscaler | AnimeSharp (4x) |
Good at: semi-realistic and photoreal-leaning renders in a family that is otherwise all
anime, the widest CFG range on the page (2 to 10), and reproducible seeds -- dpmpp_2m is
deterministic, so re-running a seed gives you the same image. It writes the score_9,
score_8_up, score_7_up prefix for you, and its AI enhance template
does the same.
Weak at: flat cel-shaded anime, which is what everything else here does better. Pony's tag vocabulary predates Illustrious and NoobAI, so newer character and artist tags land less reliably. It has the highest step floor in the family (30 minimum, up to 40), so it is never a fast pick. And because it is the only Pony model, there is no alternative if the look is wrong -- you have to change subtype entirely.
Notes: the deep_negative_pony embedding targets this subtype. Character matching still
writes Danbooru-escaped names here, the same as on Illustrious and NoobAI.
Resolutions in full#
You can pick an aspect ratio instead of a size. Each subtype bucket has its own snapped table.
Illustrious bucket (~1.77 MP, default 1152x1536):
| 1:2 | 9:21 | 9:16 | 2:3 | 3:4 | 4:5 | 7:9 | 1:1 | 9:7 | 5:4 | 4:3 | 3:2 | 16:9 | 21:9 | 2:1 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1024x2048 | 768x1792 | 864x1536 | 1024x1536 | 1152x1536 | 1216x1520 | 1184x1520 | 1536x1536 | 1520x1184 | 1520x1216 | 1536x1152 | 1536x1024 | 1536x864 | 1792x768 | 2048x1024 |
NoobAI / Pony / plain SDXL bucket (~1.03 MP, default 896x1152):
| 1:2 | 9:21 | 9:16 | 3:5 | 2:3 | 3:4 | 7:9 | 1:1 | 9:7 | 4:3 | 3:2 | 16:9 | 21:9 | 2:1 | 5:2 | 3:1 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 704x1408 | 704x1344 | 768x1344 | 768x1280 | 832x1216 | 896x1152 | 896x1152 | 1024x1024 | 1152x896 | 1152x896 | 1216x832 | 1344x768 | 1344x704 | 1408x704 | 1600x640 | 1728x576 |
You can also type a literal WIDTHxHEIGHT. It is snapped to a multiple of 8, floored at 64,
and scaled down proportionally if the longest side goes past 4096.
What works on SDXL that does not elsewhere#
| Feature | On SDXL |
|---|---|
| Negative prompt | Yes, fully honoured |
| Automatic quality tags | Yes, this is the only family that gets them |
| Detailer | Yes |
| USDU upscale | Yes -- and SDXL is the only architecture where USDU feeds your real prompt into each tile. Every other family gets a fixed generic detail prompt instead. |
| Hi-res fix | Yes, default denoise 0.35 |
| img2img | Yes. Your input image is normalised to about 1 MP before it is encoded, whatever size you uploaded. |
| Embeddings | Yes, this family only |
Warning
Some LoRAs override your sampler and scheduler for the whole job. The first LoRA in your stack that declares one wins, silently, and the change shows up in the 🔍 Details view rather than on the card. On the two v-pred models that is worth checking before you blame the prompt.