Translated from the original Japanese article
Tech 7 min read

Anima checkpoints: 20+ derivatives sorted by type and style, one LoRA fits all

AI Image Generation ComfyUI Qwen Stable Diffusion LoRA Anima WAI-Anima
IkesanContents

CivitAI now has an “Anima Checkpoint” category, and the number of Anima-derived checkpoints has exploded. Some people run the official base raw, some go the opposite way and strip quality tags, some are merges that add chibi characters, some are distilled for speed — the types and goals are all over the place, and it’s hard to tell which one to try.

So I gathered 20-plus of the main Anima derivatives and sorted them by type, aesthetic aim, and prompt adherence. This isn’t a “which one is best” verdict — it’s a survey before actually running them with my own LoRA.

Derivatives all take the same LoRA

As covered in my February post on Anima, Anima isn’t an SDXL fine-tune — it’s a DiT built on NVIDIA Cosmos-Predict2-2B, with Qwen3 0.6B as the text encoder and Qwen Image VAE. It’s a different lineage from SDXL/Illustrious, so an SDXL LoRA not loading on an Anima model is a given.

The part that matters is the reverse: every Anima derivative shares this same base, text encoder, and VAE. So a character LoRA trained for Anima loads on any of them as-is. The only reason I can run the same LoRA and the same prompt through a dozen derivatives and look only at the output differences is that the whole family sits on one base.

That said, “it loads” and “the character comes out at the same quality” are two different things. In my Anima-Base v1.0 vs WAI-Anima test, the same kanachan LoRA tended to fall apart on the raw base-v1.0, while WAI-Anima — with its baked-in aesthetic tuning — pulled the character out more reliably. Whether a LoRA loads is decided by the architecture; how faithfully it comes out depends on each derivative’s tuning.

graph TD
  C[NVIDIA Cosmos-Predict2-2B] --> A[Anima base official]
  A --> P[preview3-base older dev]
  A --> W[WAI-Anima fine-tune]
  A --> FT[Fine-tunes<br/>AnimaYume / JANIMA / Kirazuri etc.]
  A --> MG[Merges<br/>AnimaIka / Nova Anime AM etc.]
  W --> MG

A few constraints apply across the board.

  • The license is CircleStone Labs’ non-commercial license. Derivatives generally inherit it, so commercial use is out.
  • ControlNet isn’t supported. External pose or composition control doesn’t work for now.
  • The text encoder and VAE are separate files from the model itself. Every derivative needs the Qwen3 TE / Qwen Image VAE placed at install time.

The Anima derivative catalog

I pulled the top of CivitAI’s “Anima Checkpoint” category by popularity and recency. Download and rating counts are rough figures as of early June 2026; CivitAI updates constantly, so read them as ballpark scale rather than ranking.

Base models (the reference points)

  • Anima-Base v1.0: CircleStone Labs’ official “true base.” No aesthetic tuning, so it puts out plain images unless you use quality and artist tags. The starting point for every derivative.
  • Anima preview / preview3-base: the pre-v1.0 dev builds. Many early merges and fine-tunes are based on these. If you’re starting now, v1.0 is the default.

The WAI line (my daily driver)

  • WAI-Anima: a derivative that WAI0731 (known for WAI-Illustrious) fine-tuned from preview3-base. The art style is polished and consistent; as in the post where I first tried it, it’s my daily driver.

Mainstream fine-tunes (faithful to LoRA and prompt)

  • AnimaYume v0.5: an Anima fine-tune that also pulls in some e621 concepts; NSFW-capable. At ~30k downloads, among the most popular derivatives.
  • JANIMA v1.0: a fine-tune of base-v1.0. Claims a design that doesn’t force a single art style, so it’s easy to pair with LoRAs.
  • Anima Cat Tower v1.0: a fine-tune of base-v1.0 that strengthens the anime style. Also frequently used as a building block in other merges like AnimaIka.
  • Animax v0.5: flat colors and sharp lines, leaning 2D anime. The art style shifts a lot with the sampler.

Large / high-variance fine-tunes

  • Kirazuri v2.0: a full fine-tune of preview3-base on 35,537 images. Centered on 1536px, with a strong house style. Short prompts tend to bleed concepts; long, natural-language prompts ease it.
  • Hassaku Anima v0.1: base-v1.0-based, leaning 2.5D, aiming for Illustrious-grade quality. But v0.1 runs a high learning rate, and depending on the character, hands and colors can break.

Contrarian designs (no quality tags, short prompts)

  • Hikari Anima 0.4.1: drops the dependence on quality tags and adds film grain to suppress the obvious “AI masterpiece” look. Since it doesn’t auto-apply quality tags, you compensate with Danbooru tags or natural language.
  • SimpleAnima v1.0: a merge built around short prompts that aims to come out clean without piling on long quality tags or negatives.

Distilled (handled differently)

  • RDBT | Anima v0.37.1: an individual fine-tune that pushed the base on high-aesthetic data and then applied guidance distillation. Runs at CFG 1–4, and with CFG off it’s about 2x faster. An explicit style is required; with none specified the style goes random. Merge redistribution is prohibited.

Merges

  • AnimaIka v4.0: a merge of AnimaIka v3.5 + Anima Base v1 + Anima Cat Tower v1 + WAI-Anima v1. It contains WAI-Anima itself, so from a WAI user’s view it’s basically “WAI plus extras.”
  • Nova Anime AM v2.0: a base-v1.0-line merge claiming finger stability, multi-subject handling, and a 2.5D/3D lean. Updated 5/31 and recent, with downloads climbing fast since release.
  • CottonAnima base1: mixes a style LoRA into the base to unify the art style. Consistency-focused, at the cost of variety.
  • CapAnima Base1: a merge that mixes several LoRAs into the base to fill in the chibi characters Anima is weak at.
  • Anime Studio V2 (Anima): an Anima-specialized, illustration-leaning merge. Stands out for a very high review count.
  • Hotaru Blend Anima: aimed at clean-lined anime, handling both tags and natural language. The latest v4 is a merge from base-v1.0.
  • Animosity: high saturation, ships with a modified VAE. The same model page lists an Anima version and an Illustrious version as separate checkpoints (it isn’t mixing DiT and UNet — they’re separate by lineage). Use the Anima version.

NSFW-leaning and misc

  • Easily Anima: NSFW-leaning, placed on the adult-content domain. Details unconfirmed.
  • Master Anima V2: essentially a “how to make an Anima merge” tutorial. It ships with layer-wise blend scripts, for people who want to build their own merges.

The two axes the derivatives split on

Lay out the 20-plus and they scatter along roughly two axes.

The first is “how faithful to the LoRA and prompt” vs “how much art style the model carries on its own.” The raw base, JANIMA, Hikari, SimpleAnima, and RDBT (which assumes an explicit style) lean faithful — easy to steer toward your own LoRA or prompt. WAI-Anima, the various merges (AnimaIka, Nova, CottonAnima), and the strongly-styled Kirazuri push the model’s own look harder.

The second is “quality tags assumed” vs “designed to not need them.” The base and Kirazuri assume you’ll stack quality and artist tags; Hikari and SimpleAnima aim to come out clean even without them.

This isn’t about better or worse — it’s the axis you choose by how you want your LoRA to sit. If character fidelity is the priority, lean to the low-style side; if you want output that’s usable out of the box, the pre-tuned merge side.