DEV Community

AI-KSK
AI-KSK

Posted on Edited on Originally published at huggingface.co

Anima 2B: A Small Anime Model Trained Without AI-Synthesised Data

Most anime image models are trained on datasets that include AI-generated images. Anima 2B is notable for the opposite: no AI-synthesised data in the training set. That is a claim worth understanding, because it predicts something specific about how the model behaves.

The loader stack

Node Value
UNETLoader anima-preview.safetensors
CLIPLoader qwen_3_06b_base.safetensors, type stable_diffusion
VAELoader qwen_image_vae.safetensors

At 2B parameters this is a small model — roughly a tenth of what the larger anime bases carry. That constrains what it can hold, and the release is honest about being a preview.

What "no AI data" predicts

If a model never saw AI-generated images, two consequences follow:

Fewer amplifier artifacts. Models trained on AI output learn the artifacts of other AI models — over-smooth skin, characteristic eye rendering, a particular quality of gradient. A model without that training does not reproduce those artifacts.

Less "AI look". The default output style is closer to human illustration, because the training distribution was human illustration.

Both are testable: generate the same prompt on this model and on a model known to include AI data, and look at whether the surface treatment has that recognisable sheen.

Where a 2B model loses

Parameter count is a hard limit on what a model can represent. At 2B:

  • Prompt complexity — long prompts with many constraints degrade. Fewer "slots" to hold them.
  • Fine detail consistency — hands, small objects, symmetrical features.
  • Style range — fewer distinct artist styles are well-represented.

None of that is a flaw; it is the trade for speed and low memory.

The Gundam Flux LoRA is still there

The block upscale graph loads:

LoraLoader → Gundam_RX78_Flux.safetensors at weight 0.85

Same finding as in multiple other Anima releases: a Flux-architecture LoRA inside an Anima model. Not architecturally valid, loads without error, produces a styled result that you cannot attribute.

This LoRA has now appeared in at least four Anima releases at the same 0.85 weight in the same block-upscale graph. It is a copying-forward artifact. Clear that slot.

Where this release is weak

  1. The Flux LoRA.
  2. 13-node image-to-image has minimal validation.
  3. 2B capacity limits prompt complexity — if your workflow depends on long structured prompts, this is the wrong model.
  4. Preview, not final. anima-preview is an early checkpoint; later Anima versions differ.

Practical approach

Use this model when you want fast iteration at low memory and your prompts are short and tag-style. Use a larger anime base when you need complex prompts or fine detail.

If you want to test the "no AI data" claim, generate a portrait with neutral lighting and compare surface treatment against a model you know included AI data. The difference, if present, shows in the skin and the eyes.

Workflows

Three graphs, published on RunningHub with invite code rh-v1111:

  • Anima text-to-image
  • Anima image-to-image (anime conversion)
  • Anima block/tiled upscale

Files

https://huggingface.co/AI-KSK/ksk-video-177-comfyui-workflows

Notes

  • Workflow files only. Model weights are not included; each keeps its own licence.
  • Tutorial exports, not a product.
  • The "no AI-synthesised data" property is the release's stated claim about training data, not something these workflows can verify.
  • Independent AI-KSK release. Not affiliated with the model authors or RunningHub.

Optional API platform - 88API (affiliate link)

Explore 88API for access to multiple AI model APIs. Available models, routes and prices are listed on the platform; Alipay top-ups are supported.

https://88api.ai/sign-up?aff=usx3

Disclosure: I may earn a commission from eligible purchases after you register through this referral link.

Top comments (0)