# Released Wulver v0.5, our Krea 2 finetune for anime and furry. — retained record

> Retained reddit record from reddit:StableDiffusion.

Canonical: https://highsignal.app/data/records/285b9cd1fb449b82

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:StableDiffusion)  published 2026-09-30 14:54 ingested 2026-10-01 02:41

[original source ↗](https://www.reddit.com/r/StableDiffusion/comments/1wu721p/released_wulver_v05_our_krea_2_finetune_for_anime/)

## retained text



```
Wulver v0.5 is a full fine-tune of Krea 2 Raw (the whole 12.8B DiT) for anime and anthro

characters, including scenes where several characters interact. We made it at Vaelico,

an independent studio.



Compared with v0.1, which we posted here earlier, v0.5 comes from a far longer and more

serious training run. We tested the two side by side on about 1,000 test images covering

every artist on the list and a very wide range of prompts, and v0.5 is clearly better in

most regards. If you kept your v0.1 prompts, run them again.



\`@artistname\` works now. v0.1 mostly ignored the prefix; v0.5 has 1,113 artist styles that

respond to it, and the full list is in [ARTISTS.md](http://ARTISTS.md) on Hugging Face.



It rewards detailed, descriptive prose in the prompt, and it also takes booru-style tag

lists. For artist styles, keep the prompt short for now: long prompts dilute them.



\- Hugging Face: [https://huggingface.co/Vaelico/Wulver](https://huggingface.co/Vaelico/Wulver)

\- Civitai: [https://civitai.com/models/2881657](https://civitai.com/models/2881657)

\- No GPU? Run it in the browser, with free images every day: [https://vaelico.ai/?src=reddit](https://vaelico.ai/?src=reddit)



For launch week, free accounts get 100 images a day (1024 px) instead of the usual 15,

until October 7, 23:59 UTC.



\*\*Turbo\*\* (official Krea 2 Turbo LoRA merged in; 8 steps, CFG 1, euler/simple, up to 2K):



\- fp8 e4m3fn, 12.8 GB: the one most ComfyUI users want

\- int8 convrot, 13.5 GB: for forge-neo and other int8 runtimes

\- w4a8 convrot, 7.7 GB: the smallest, needs ComfyUI 0.31 or newer

\- bf16



\*\*Non-Turbo\*\*: bf16 for LoRA training, further fine-tuning or setting your own turbo

strength, plus an int8 convrot.



GGUF Q8\_CR, Q5\_0 and Q4\_0 are on Civitai.



The drag-and-drop ComfyUI workflow uses stock loaders. The text encoder is Qwen3-VL-4B and

the VAE is the Qwen Image VAE, both from Comfy-Org/Krea-2, so if you already run Krea 2 in

ComfyUI, the Wulver checkpoint is the only new download.



\*\*Known issues\*\* (the first two are addressed in v1):



\- A thin smeared strip can appear at the top and bottom edges. Generate a bit larger and crop.

\- Some concepts bleed into each other or depend on the \`@artist\` you use, a side effect of

  teaching it 1,113 artist styles.

\- Some style LoRAs trained on Krea 2 Raw don't carry over (character LoRAs do).



License: Krea 2 Community License. Wulver is a modified Krea 2 model, and we're not

affiliated with Krea.



v1 is planned at a different scale: a much larger dataset, longer training at high

resolution and a dedicated post-training stage. We want partners in it from the start: a

compute provider, and companies whose products are built around anime and anthro

characters. If you can back v1 with compute or funding, write to [hello@vaelico.ai](mailto:hello@vaelico.ai) and

we'll walk you through the plan.


```



document fetched 2026-10-01 00:35 · canonical https://reddit.com/r/StableDiffusion/comments/1wu721p/released_wulver_v05_our_krea_2_finetune_for_anime
