Skip to content

Standalone Models#

Seven checkpoints that sit outside the big families: Mage Flow (2), ERNIE (2), Qwen, Mugen Aesthetic and NetaYume AIO. Each is its own architecture with its own habits, and each is the only thing in the catalog that behaves the way it does.

For the bigger families see Krea 2, Z-Image, SDXL, Flux, Ideogram 4 and Anima.

At a glance#

Model Picker entry Steps CFG Sampler Scheduler Steps range CFG range Base size Batch Negative Enhance default
Mage Flow (Quality) Mage Flow > Quality 20 5 euler simple 12-30 3-7 1024x1024 2 Yes On
Mage Flow Turbo Mage Flow > Turbo 4 1 euler simple 2-8 1-2 1024x1024 2 Yes On
ERNIE Image ERNIE > Image 20 4 euler simple 15-50 3-6 1024x1024 2 No On
ERNIE Turbo ERNIE > Turbo 8 1 euler simple 4-12 1-2 1024x1024 2 No On
Qwen direct pick 20 2.5 euler simple 10-30 1-5 1328x1328 2 Yes Off (opt-in)
Mugen Aesthetic direct pick 28 5 dpmpp_2m_sde simple 20-30 4-7 896x1152 2 Yes Off (opt-in)
NetaYume AIO (v3.5) direct pick 40 4 res_multistep linear_quadratic 35-55 3.5-5 864x1152 2 Yes Off (opt-in)

None of the seven pins a default upscaler, so hi-res fix and /upscale fall back to the global 4x-AnimeSharp.pth.

Prompt enhancement is already on for four of these

ERNIE Image, ERNIE Turbo, Mage Flow (Quality) and Mage Flow Turbo all default enhance: to ON. Every other model in this list is opt-in. Pass enhance: False if you want your literal text encoded. (Ideogram 4 is the third family that defaults on.) See AI Enhance.

None of these has a LoRA of its own

Of the 91 LoRAs in the library, not one declares compatibility with mage_flow, ernie_image, qwen_image, mugen or neta-lumina. On all seven models the LoRA picker offers only the universal LoRAs -- the ones with no compatibility list at all. See the LoRA catalog.

The models#

Mage Flow (Quality)#

Microsoft's 4B NR-MMDiT in its RL-aligned build, quantized to int8. The pick when you want a modern DiT that still listens to a negative prompt.
Architecture Mage-Flow 4B NR-MMDiT (int8)
Subtype mage_flow, mage_rl
Prompting Prose. Enhancement is on by default, so short input is fine
Negatives Supported -- one of the few DiT models here that honours them
Steps 20 (12-30)
CFG 5 (3-7)
Sampler euler
Scheduler simple
Base resolution 1024x1024
Default batch 2
Default upscaler None pinned -- falls back to 4x-AnimeSharp.pth

Good at: prose scene prompts with real CFG response from 3 to 7, actually excluding things via negative:, any aspect ratio between 512 and 2048 per side, and permissive licensing -- MIT, the most permissive model in the catalog.

Weak at: anime and character work, where the SDXL family and Anima have far more knowledge baked in; style LoRAs, since none exist for it; speed, at 20 steps and CFG 5; and it gives you the enhancer whether you wanted it or not unless you remember enhance: False.

Mage Flow Turbo#

The same architecture on a 4-step CFG 1 recipe. A genuinely different checkpoint, not a preset.
Architecture Mage-Flow-Turbo 4B NR-MMDiT (int8)
Subtype mage_flow, mage_turbo
Prompting Prose. Enhancement is on by default
Negatives Supported
Steps 4 (2-8)
CFG 1 (1-2)
Sampler euler
Scheduler simple
Base resolution 1024x1024
Default batch 2
Default upscaler None pinned -- falls back to 4x-AnimeSharp.pth

Good at: drafting a Mage Flow composition in a fraction of the time, the same resolution freedom and MIT licence as Quality, quick batches of 2.

Weak at: obeying a long prompt -- it is distilled, so the recommended CFG ceiling is 2 and turning guidance up is not an option; fine detail at 4 steps; and while the negative prompt is wired in, at CFG 1 it has very little to push against, so do not expect much from it here.

Mage Flow dimensions are strict

Both sides must be a multiple of 16 and between 512 and 2048. The shipped aspect-ratio table is already aligned, so this only matters if you pass an explicit WxH.

ERNIE Image#

Baidu's ERNIE Image, base build. 20 steps at CFG 4.
Architecture ERNIE Image (Baidu), Flux2-family VAE and text encoder
Subtype ernie_image, ernie_base
Prompting Prose. Enhancement is on by default
Negatives Dropped -- the graph has a negative slot but it is never filled from your input
Steps 20 (15-50)
CFG 4 (3-6)
Sampler euler
Scheduler simple
Base resolution 1024x1024
Default batch 2
Default upscaler None pinned -- falls back to 4x-AnimeSharp.pth

Good at: general prose-prompt generation with an unusually wide usable step range (15 to 50), Apache 2.0 licensing, and a different look from the Flux and Z-Image models it shares a VAE lineage with.

Weak at: half your input, because negative: is silently discarded -- everything must be phrased positively; control generally, since enhancement is also on by default; anime and character knowledge; and LoRAs, of which it has none of its own.

ERNIE ignores the negative prompt

A negative branch exists inside the ERNIE graph but it is never filled in from your input -- it stays an empty string. Whatever you type in negative: is dropped without a message on the card. If you need something excluded, use Mage Flow (Quality) or an SDXL model instead.

ERNIE Turbo#

The 8-step CFG 1 distill of the same model. Fourth in the default picker order, so it is one of the first things a new user sees.
Architecture ERNIE Image Turbo (Baidu), distilled
Subtype ernie_image, ernie_turbo, ernie_distilled
Prompting Prose. Enhancement is on by default
Negatives Dropped, same as ERNIE Image
Steps 8 (4-12)
CFG 1 (1-2)
Sampler euler
Scheduler simple
Base resolution 1024x1024
Default batch 2
Default upscaler None pinned -- falls back to 4x-AnimeSharp.pth

Good at: fast drafts, cheap batches of 2, and being near the top of the picker when you want something quick without scrolling.

Weak at: everything the base ERNIE is weak at, plus a hard CFG ceiling of 2 and 8 steps of detail. It is the same model with less time spent on it -- if the composition is right but the render is rough, re-run it on ERNIE Image.

Qwen#

Qwen Image 2512 in a 4-bit Nunchaku build. One of the two models in the catalog picked for readable text in the image, and the only one of those that also takes a negative prompt.
Architecture Qwen Image 2512, Nunchaku INT4 quantization
Subtype none
Prompting Prose. Enhancement is opt-in
Negatives Supported
Steps 20 (10-30)
CFG 2.5 (1-5)
Sampler euler
Scheduler simple
Base resolution 1328x1328 -- the largest square base in the catalog
Default batch 2
Default upscaler None pinned -- falls back to 4x-AnimeSharp.pth

Good at: rendering legible text without the forced-enhancement tax Ideogram 4 charges for it, large square renders, img2img (default denoise 0.7), and honouring both a positive and a negative prompt.

Weak at: fine detail relative to a full-precision model -- this is a 4-bit quantization, chosen for speed and VRAM; LoRAs, which are model-side only (the text encoder is never patched) and limited to 4, with none tagged for it; a narrow aspect-ratio table of only 7 ratios; and /turbo, which throws your LoRAs away entirely.

Qwen LoRAs are dropped on /turbo

The Qwen MrFlow preset in /turbo silently discards any LoRA you select. It is logged as a warning, not shown on the card. Use /imagine if you need them.

Mugen Aesthetic#

NoobAI-lineage SDXL weights paired with a Flux 2 VAE and flow-based sampling. Tag prompts, painterly output.
Architecture SDXL weights (NoobAI lineage) with a Flux 2 VAE, flow-matching sampling
Subtype noob, flux2vae
Prompting Danbooru-style tags, like the SDXL family
Negatives Supported, and a defect list is baked in for you
Steps 28 (20-30)
CFG 5 (4-7)
Sampler dpmpp_2m_sde
Scheduler simple
Base resolution 896x1152 (the NoobAI subtype table)
Default batch 2
Default upscaler None pinned -- falls back to 4x-AnimeSharp.pth

Good at: painterly anime with cleaner colour than plain NoobAI, tag prompts with character knowledge from the NoobAI lineage, and a working negative prompt with sensible defaults already in it.

Weak at: memory and speed -- the Flux 2 VAE halves spatial dimensions on decode, so the latent is rendered at double size and decoded in tiles, which costs more than a plain SDXL model at the same output size; safe-content control, because it ships no SFW guard strings, so your nsfw setting changes nothing in the prompt on this model; LoRAs, since the SDXL-tagged ones do not match it and it has none of its own; and prose prompts, which it handles worse than tags.

Notes: baked-in positive is masterpiece, best quality. Baked-in negative is (worst quality:1.1), normal quality, (bad anatomy:1.1), (blurry:1.1), watermark, sepia, (adversarial noise:1.1), jpeg artifacts. If you pass raw: True, both disappear along with everything else.

NetaYume AIO (v3.5)#

A Neta-Lumina all-in-one anime checkpoint. The slowest model on this page and the one with the strangest prompt handling.
Architecture Neta-Lumina (all-in-one checkpoint)
Subtype none
Prompting Prose, wrapped in an instruction sentence automatically
Negatives Supported, on top of a long baked-in defect list
Steps 40 (35-55)
CFG 4 (3.5-5)
Sampler res_multistep
Scheduler linear_quadratic
Base resolution 864x1152 (the global default -- it ships no table of its own)
Default batch 2
Default upscaler None pinned -- falls back to 4x-AnimeSharp.pth

Good at: long natural-language anime descriptions, a strong built-in negative that catches the usual hand and anatomy failures without you writing one, and a look distinct from every SDXL model here.

Weak at: turnaround -- 40 steps at a default batch of 2 is the longest wait on this page; resolution, since it has no per-model aspect table and falls back to the general 864x1152 budget; post-processing, because both the Detailer and USDU refuse it outright; and LoRAs, of which it has none.

Your prompt is wrapped, not sent

NetaYume prepends You are an assistant designed to generate high quality anime images based on textual prompts. <Prompt Start> to your positive prompt, and a matching low-quality instruction plus a long defect list to your negative. That happens automatically -- do not type it yourself. The AI enhancer treats this model like Z-Image, producing a long natural-language paragraph.

Detailer and Ultimate SD Upscale refuse to run

Both reject NetaYume with an error. There is no face, eye or hand fix pass available for a NetaYume image -- if the anatomy comes out wrong you re-roll it. For enlarging, use the Model-based (fast), Refine or SeedVR2 methods on /upscale. See the Detailer and Upscaling methods.

Resolutions#

Mage Flow and ERNIE share one table, default 1024x1024 -- the same one Klein uses:

1:1 9:7 7:9 4:3 3:4 3:2 2:3 16:9 9:16 21:9 9:21
1024x1024 1136x880 880x1136 1152x896 896x1152 1216x832 832x1216 1344x768 768x1344 1536x656 656x1536

Qwen, default 1328x1328, with only seven ratios:

1:1 16:9 9:16 4:3 3:4 3:2 2:3
1328x1328 1664x928 928x1664 1472x1104 1104x1472 1584x1056 1056x1584

Mugen Aesthetic uses the NoobAI subtype table (base 896x1152, 1:1 is 1024x1024). NetaYume AIO ships no table at all and falls back to the global 864x1152 budget, with other ratios computed from it.

Editing#

Mage Flow is also an editing model, with fast (4 steps, CFG 1) and quality (20 steps, CFG 5) modes, up to 3 reference images, and the first image normalised to about 1 MP so the output size stays sane whatever you upload. Batch is forced to 1. See /edit.

Qwen supports img2img through /imagine with an attachment -- your input is normalised to about 1 MP and the denoise defaults to 0.7.

The other five on this page are text-to-image only.

Categories: Models | Prompting | Imagine | AI Enhance