Celevi is a fan-made, anthropomorphic take on Celebi, the mythical Grass-type Pokémon, reimagined as an adult woman in her twenties and drawn as a polished anime illustration. It is a safe-for-work character LoRA: it was made and tested for clothed, everyday and fantasy scenes, and every image in the gallery is PG or PG-13.
Her design
Pale mint-green skin and a round face with large, expressive blue-green eyes.
A leaf-shaped crest swept back over her head, and two long antennae.
Large, translucent, iridescent wings rooted in her upper back.
A simple mouth: closed or a little "o", never teeth.
Her signature outfit (
tirvel outfit, leaf clothing): layered emerald leaf plates forming a strapless one-piece, with leaf gloves to just below the elbow; legs and feet bare.
Quick start
Pick the version for your checkpoint's family (list below).
Start with the trigger
celevi, and describe her as a woman.Name an outfit in every prompt:
in her tirvel outfit, leaf clothingfor the signature look, or any other outfit in plain words (wearing a long red evening gown,wearing a fitted cream knit sweater).LoRA weight 0.8 to 1.0 (SDXL: up to 1.2).
On fine-tuned checkpoints (aMix, WAI, AutismMix), begin the prompt with
green skin, colored skin, or her skin drifts toward peach.Want the wings? Ask for them:
wings spread behind her(MOSS-Image often leaves them out otherwise).
Which version?
Illustrious: for Illustrious XL and its fine-tunes (the gallery uses aMix). The most polished look; follows prompts the least (44%).
Pony: for Pony Diffusion V6 XL and its fine-tunes (the gallery uses AutismMix). The strongest likeness (82% on her design checklist); needs Pony's score tags.
SDXL: for SDXL 1.0 base. Follows prompts well (76%); fine-tunes that retrained the text encoders (Animagine XL) lose her.
MOSS-Image: for MOSS-Image SFT and Base, in ComfyUI. Follows prompts best (88%); ask for the wings.
The percentages come from 17 evaluation images per version, scored against her design checklist (details in each version's notes).
What works well
Medium and full shots: walking through a town, reading in a library, hiking a mountain path, a greenhouse, a forest at night, an evening party.
Other outfits in plain words: long dresses and evening gowns, knitwear, a travelling cloak, a sun hat and a maxi dress by the sea.
Calm, warm and confident expressions.
Known limits
Profile and back views are less reliable than front and three-quarter views.
Extreme close-ups with strong expressions can drift toward a generic face; medium and full shots hold her design best.
The signature outfit often gains extra pieces, such as leaf boots or longer gloves.
The SDXL and MOSS-Image versions are a step behind Illustrious and Pony; updates are planned.
How it was made
125 AI-generated reference images of her (front, three-quarter, profile and back views; several outfits and expressions), each caption checked by hand.
One LoRA per base model, trained on the same images and captions; each version's checkpoint was picked by the creator in side-by-side comparisons.
The gallery was rendered in ComfyUI with each version's recommended settings plus a 1.5x hires pass (MOSS-Image: plain text-to-image). The generation data is in every image.
Feedback
Share what you make with her, and leave a review or a comment on what works and what doesn't: it shapes the next version.
Fan work: not affiliated with or endorsed by Nintendo, Creatures, GAME FREAK or The Pokémon Company, who own Pokémon.
Trigger word:
celevi· Outfit tag:tirvel outfit, leaf clothingShe is an adult woman, in her twenties. Prompt her as one, in adult settings and poses, and name an outfit in every prompt.
Description
Trained on [Anima Base v1.0](https://huggingface.co/circlestone-labs/Anima) anima-base-v1.0), and tested on that checkpoint.
## Read this first: Anima
Anima is not an SDXL model: it is CircleStone Labs' diffusion transformer, with its own text encoder and VAE. Use this LoRA with Anima's files, in ComfyUI or another UI that supports Anima:
- ComfyUI: load anima-base-v1.0.safetensors with the diffusion-model loader, the text encoder qwen_3_06b_base.safetensors (type stable_diffusion) and the VAE qwen_image_vae.safetensors, all from Anima's page; then this LoRA with a LoRA loader on the model.
- Trained on Anima-Base, as Anima's authors advise for LoRAs. Their Aesthetic and Turbo versions may work but are untested.
## How to prompt
Anima reads tags, natural language, or a mix. Quality and rating tags first, then the trigger:
```
masterpiece, best quality, safe, celevi, a young woman, standing in a sunlit forest clearing, gentle smile, looking at the viewer, three-quarter body framing, in her tirvel outfit, leaf clothing, refined anime illustration
```
- Quality tags: masterpiece, best quality, safe safe is Anima's rating tag: keep it for safe-for-work images)
- LoRA weight: 1, the weight this version was tested at.
## Recommended settings
- Sampler: Euler (ComfyUI: euler, scheduler simple)
- Steps: 30 (the base model's range: 30 to 50)
- CFG: 4 (the base model's range: 4 to 5)
- Size: 832x1216 (Anima works from about 512x512 to 1536x1536)
Showcase images: Anima Base v1.0 at weight 1.0, green skin, colored skin right after the trigger and class phrase, 832x1216 (one scene square, one landscape), Euler, 30 steps, CFG 4, in ComfyUI; no hires pass, no face detailer.
## Negative prompt
```
lowres, worst quality, bad anatomy, watermark, jpeg artifacts, extra limbs, nsfw, nude, naked, nipples, exposed breasts, bare buttocks, teeth, clenched teeth, fang, fangs
```
The first words are the usual quality negatives. The rest keep her clothed (this is a safe-for-work model) and keep her mouth as designed: closed or a small "o", never teeth.
## How it was checked
The checkpoint at step 600 was picked by the creator on side-by-side sheets: this gallery's 16 prompts, rendered at five checkpoints (steps 200 to 1000). Like the NoobAI V-Pred version, this one has no design-checklist scores.
## Known limits
- Tested on Anima Base v1.0 itself, at weight 1.0 with Euler and 30 steps; Anima's Aesthetic and Turbo versions, its fine-tunes, other weights and other samplers are untested.
- No design-checklist scores yet: the gallery is this version's evidence.
## Training
- LoRA rank 32, alpha 32; 600 steps (9 epochs); learning rate 5e-05 (cosine); AdamW8bit; trained on Anima-Base v1.0.
- Dataset: 125 AI-generated images of the character (front, three-quarter, profile and back views; several outfits and expressions), each with a hand-approved caption. The caption phrases and their counts are in the file's metadata (ss_tag_frequency); the images themselves are not published.
## Licence
This LoRA is a derivative of [Anima](https://huggingface.co/circlestone-labs/Anima) (CircleStone Labs; a derivative of NVIDIA Cosmos-Predict2-2B, so NVIDIA's Open Model License applies too), modified by training this LoRA on Anima-Base, and it is released under the base model's licence: [CircleStone Labs Non-Commercial License v1.2](https://huggingface.co/circlestone-labs/Anima/blob/main/LICENSE.md).
The CircleStone Model is licensed by CircleStone Labs LLC under the CircleStone Non-Commercial License.
The licence is non-commercial for the model and this LoRA; it does not restrict the images you make (its section 2(e)). It is a free download; the training recipe is above.
## Changelog
v1.0
- First release of the Anima version: dataset v0.2 (125 training images), checkpoint at step 600.















