.ai

de bästa AI-modellerna med öppen källkod

Sök

Anima: light, fast, slightly unruly

Q: Over-saturated colours, harsh outlines, burned highlights

CFG pushed too high; above about 10 Anima burns out. Bring CFG back to 4 to 5.

Q: Text in the image is garbled

Text rendering is at the weak end by today's standards. Keep on-image text to one or two words or add it in editing.

Ämne: Modeller

Av Chris Green

Publicerad 2026-09-26

An anime, manga and illustration model that is the antithesis of Z-Image Turbo: 2 billion parameters, 4 GB, fast on modest GPUs, natural language and Danbooru tags in the same prompt, and happiest when given room to be creative. By Chris Green of Diffusion Doodles.

Den här sidan har inte översatts ännu, så den visas på engelska.

Senaste informationen

Den här guiden är en ögonblicksbild. Källan nedan hålls uppdaterad av sina skapare; se där för de senaste detaljerna.

Diffusion Doodles, Chris Green's Substack

Översikt

Anima

is a text-to-image model made by

ComfyUI

together with CircleStone Labs, an AI start-up. It appeared through a few preview releases and a v1 base model without much fanfare. The headline is that it is not a photorealistic model: the release notes say it was trained on several million anime images and around 800 thousand non-anime artistic images, with no synthetic data, and that the anime training data stops in September 2025.

The result is an open-source, compact, 2 billion parameter model built from the ground up for anime, manga, illustration and artistic concept art, an area that

SDXL

and its fine-tunes such as

Illustrious

and Pony have dominated. Built for that purpose it produces cleaner linework, better anime anatomy, more consistent character rendering and less accidental realism.

Against SDXL: SDXL is an older UNet with CLIP text encoders; Anima is a lightweight transformer built on a streamlined version of NVIDIA's Cosmos architecture. Anima reads prompts through a

Qwen

language adapter, so it understands natural language and follows prompts well, while still supporting Danbooru tags. Its 16-channel VAE (SDXL has 4) gives better lighting, composition and background detail. Being a flow-matching model it is slower than highly optimised SDXL checkpoints on consumer GPUs, but still very fast next to heavy photorealistic models such as Qwen and

Flux

. SDXL has the huge ecosystem of fine-tunes and

LoRAs

and SDXL LoRAs are not compatible, though new Anima LoRAs appear all the time. SDXL's base licence is commercially permissive; Anima's is more restrictive and geared to personal hobby projects.

In use: text rendering is weak by today's standards, best kept to one or two words. Prompt adherence is good but not at

Qwen Image

Flux.2

level; Anima almost rebels against prompts that are too restrictive and works far better with a loose guide and room to be creative. It is strongest at anime, but with the right prompt and tags it also does watercolour, sketches, 3D, oil and quite abstract output, and LoRAs take it further. Expect to vary prompt, seed and

sampler

over a few generations before you hit the right one. With the base model only a couple of weeks old, fine-tunes and LoRAs are expected to multiply over the coming months.

Referens

Namn	Typ	Vad det är
Model	file	The v1 base model (and the earlier previews) from Hugging Face, only 4 GB.
Text encoder	file	The 0.6B parameter Qwen3 text encoder (1.2 GB), from the same Hugging Face page.
VAE	file	The Qwen Image VAE from Hugging Face; you already have it if you use the Qwen Image models.
VRAM	hardware	Officially 8 GB; users report 6 GB works. On a 16 GB card it typically uses only about half.
Resolution	setting	The official documentation suggests 1 to 2 megapixels. 1:1, 3:4, 4:5 and 16:9 all work; up to 1536x1536 runs fine but quality is slightly better and more consistent a little lower.
Steps	setting	30 to 50 suggested for the base model; 40 or above gives the best quality, 30 can still be good depending on the sampler. An early Turbo LoRA brings it down to 8 to 12 steps, with some odd compression of limb lengths in places.
CFG	setting	4 to 5. Anima gives cleaner results with lower guidance; above about 10 the colours over-saturate with harsh outlines and burned highlights.
Sampler / scheduler	setting	er_sde with simple is a good default (crisp line art, flat anime colouring, consistent). Euler A: softer lines, more painterly, slightly 2.5D, good for character art and covers. DPM++ 2M SDE GPU: more variation and richer compositions, unpredictable with complex prompts. The standard KSampler; the ClownShark sampler has not given good results so far.
Negative prompt	prompt	Supported. Use it to steer away from styles such as "photorealistic" and for the usual tags: blurry, low resolution, oversaturated, watermark, messy typography, distorted anatomy, overexposed highlights, bad hands, malformed hands.
Safety tags	prompt	Anima is fairly uncensored and can produce undesired content from short or vague prompts. Put safe, sensitive, nsfw or explicit in the positive or negative prompt as appropriate, especially when reusing short Danbooru-style prompts.
@artist	prompt	Artist reference tags must be preceded by the @ symbol and are best placed at the start of the prompt. The Anima Style Explorer lists some 60,000 artist styles from the training set.

Inget att installera - kör AI-modeller i molnet

BitVector Prism är det enklaste sättet att komma igång: välj en modell, skriv en prompt och generera i en ren webbapp, utan någon installation alls. BitVector finns också på Discord och på webben (SpyGlass).

Prova BitVector

Steg för steg

Load the three files in ComfyUI: the Anima base model, the Qwen3 0.6B text encoder and the Qwen Image VAE. Anima is supported natively and ships with basic example templates in the latest ComfyUI; the standard workflow is very simple.
Start from the standard template: KSampler, er_sde sampler, simple scheduler, 30 to 40 steps, CFG 4 to 5, a 1 to 2 megapixel canvas.
Write the prompt in plain language, two to three short paragraphs at most. Very short prompts give the model a lot of creative scope; very long prompts degrade adherence and sometimes quality.
Mix in Danbooru-style tags where they help, with any style or @artist references at the beginning. Keep it clean and concise: Anima is sensitive to repetitive tag overloading.
Add a negative prompt with the styles you do not want (for example photorealistic) and the usual quality tags, and a safety tag when the positive prompt is short or vague.
Generate and look at the result as a draft. If it misses the detail you wanted, loosen rather than tighten the wording, then vary seed, tags and sampler over a few generations.
When you want a different look, switch the sampler first (Euler A for painterly, er_sde for crisp flat anime, DPM++ 2M SDE GPU for variety), then try a LoRA; a LoRA adds only a couple of seconds.

Exempel

A Danbooru-style prompt

1girl, solo, looking at viewer, masterpiece, detailed eyes

The usual negative prompt

photorealistic, blurry, low resolution, oversaturated, watermark, messy typography, distorted anatomy, overexposed highlights, bad hands, malformed hands

Reference settings

RTX 4060 Ti 16 GB, standard workflow, er_sde / simple, 30 steps: about 1.8 s per step, a finished image in under a minute, VRAM at roughly 50%.

Tips

Danbooru tags are token-efficient and leave the model room to be creative; natural language gives more control. Users report that hybrid prompting delivers the best balance between prompt comprehension and anime aesthetics.
Danbooru tags come from Danbooru, an anime imageboard launched in 2005 whose searchable tag system became one of the most comprehensive image taxonomies on the internet; many anime models were trained on Danbooru-tagged images, which is why tag prompting works.
Simple JSON or YAML structured prompts work because the Qwen encoder interprets them, but they break down quickly as they get complex and nothing suggests they produce better images than natural language or tags.
Switching samplers or schedulers barely changes generation time, but with Anima it can change the output a lot; the same prompt and seed through euler, euler-a, er-sde and dpmpp-sde-gpu give four noticeably different pictures.
Lower CFG than you would use on other anime models. Some of them like very high CFG; Anima does not.
If you want precise control, use a model like Flux.2. Anima is a free-flowing anime, illustration and art model that takes a rough idea and interprets it in many different ways.
Expect anime to be strongest, then watercolour, sketches, 3D and oil styles. A watercolour prompt plus a LoRA such as Vector Paintings by Daalis pushes the style further.
The Hugging Face model card has more on tag order, syntax and the natural language approach.

Keep on-image text to one or two words or add it in editing.

Inget att installera - kör AI-modeller i molnet

PirateDiffusion finns bara på Telegram och är byggt för proffs: tusentals modeller, LoRA och arbetsflöden som styrs med chattkommandon, med obegränsad generering till ett fast pris.

Prova PirateDiffusion

Frågor

Do SDXL LoRAs work with Anima?

No. Anima uses a different architecture, so SDXL LoRAs are not natively compatible. New LoRAs trained for Anima are appearing all the time.

Does Anima support JSON prompting?

Not natively as far as is known. Simple JSON or YAML prompts work because the Qwen text encoder interprets them, but they break down quickly as they get complex and do not produce better images than natural language or tags.

How fast is it?

On an RTX 4060 Ti 16 GB with the standard workflow, er_sde / simple and 30 steps, about 1.8 seconds per step and a finished image in under a minute, using about half the

VRAM

. A LoRA adds a couple of seconds.

Can I use it commercially?

Anima's licence is more restrictive than the base SDXL licence and geared toward personal hobby projects. Read the licence before commercial use.

Is it only for anime?

Anime is where it is strongest, but with the right prompt and tags it also produces watercolour, sketches, 3D, oil and quite abstract output, and LoRAs take it into further styles.