# Krea 2 - Wulver
Full fine-tune of Krea 2 Raw (12.8B DiT) — anthro & furry, with strong anime / kemono
range and true multi-character composition. Trained on 1.5M+ curated images with
multi-format captions.
> In Shetland folklore, the Wulver is a wolf-headed being that was never cruel —
> it fished the lochs and left its catch on the windowsills of those in need.
This is a v0.1 *beta** — expect rough edges. An aesthetic pass is in active development
and updates will roll out as the base training matures.*
## What it does
- Anthro / furry as its home turf — species knowledge far beyond generalist models
- Anime & kemono styles — not the usual western-model anime approximation
- Multi-character scenes — characters interacting, not merged into one blob
- Artist styles via @artistname prefixes (Curently under development)
- Best results with detailed descriptive prose — see prompting tips below
## Settings
| | |
|---|---|
| Steps | 14 |
| CFG | 1.5 |
| Sampler / Scheduler | euler / simple |
| Resolution | 1024 native |
| CLIPLoader type | krea2 |
## ComfyUI setup
1. Wulver_v0.1_fp8.safetensors → ComfyUI/models/diffusion_models/
2. Text encoder qwen3vl_4b_bf16.safetensors from [Comfy-Org/Krea-2](https://huggingface.co/Comfy-Org/Krea-2) → models/text_encoders/
3. VAE qwen_image_vae.safetensors from the same repo → models/vae/
4. Drag the attached workflow into ComfyUI and hit queue.
## Prompting tips
This model was trained on long natural-language captions, so detailed prose gets you
the most out of it. Template that works:
> *A digital painting of [subject & species, appearance details]. [What they are doing,
> pose, expression]. [Clothing / body details]. The background features [setting,
> lighting, atmosphere]. [Artistic style / medium notes].*
- 60–120 words is the sweet spot. Short prompts work, detail works better
- @artistname, at the very start for artist styles
- Keep CFG at 1.0 — this is a guidance-distilled blend; higher CFG burns the image
- Copy the prompts from the sample gallery to see the dialect in action
## Compatibility notes
- LoRAs trained on base Krea 2 may behave differently on Wulver (weights have moved)
- bf16 version available on request
## License
This is a modified version of the Krea 2 model by Krea.
Krea 2 is licensed under the [Krea 2 Community License Agreement](https://krea.ai/krea-2-licensing) —
by using this model you agree to its terms (including [Acceptable Use Policy])(https://www.krea.ai/krea-2-use-policy).
Not affiliated with or endorsed by Krea.
---
by Vaelico — generate online shortly at [vaelico.ai](https://vaelico.ai)
Description
FAQ
Comments (23)
need ... int8convrot version ... so badly ...
Is up now. good call btw, it came out basically identical to bf16
@VaelicoAi @VaelicoAi can I ask, which is the int8 version? or is it not on civ?
@sololemon Uploading the int8 safetensors right now, convrot also.
Wulver_v0.1_int8.safetensors
Wulver_v0.1_int8_convrot.safetensors
Is this model available on Huggingface as well?
https://huggingface.co/Vaelico/Wulver
Now it is :)
@VaelicoAi Can you upload it to tensor.art? (*/ω＼*)
@sky_bai I was checking that. Will take a look soon. Will ping youif you want.
Is it trained only on furry dataset or other normal anime characters also?
It is trained in normal anime characters too.
@VaelicoAi Oh it couldn't recognize Reze from Chainsawman I thought since she is fairly known character it would pick up, maybe I prompted wrong.
@AR7FX
then i will try to put more distribution like that on next. Thanks for the idea.
Right now it was mainly how god it looked, not distribution.
Hi, question: It need a turbo lora?
oh nice, a specialized anime model with a big text encoder! excited to see how this goes
Does a pretty good job with realism/semi-realism, especially when paired with the Lenovo UltraReal LoRA. Has some issues with realistic NSFW stuff, though (mainly with genitals, penetration, etc). I hope future versions improve the realism, but I know that's not really the focus of this checkpoint. Overall, great work!
It's a really fucking good alpha model!
So much potential in this one.
There is some weirdness when it comes to describing styles, often it just makes them randomly regardless of the prompt, but in any way the model itself is actually really solid.
Aside from artist tags, I hope it's going to know more custom species like Synth or Sergal (maybe even tetton) in future versions.
Nice work!
a question out of curiosity: is your dataset mainly with e621 or furaffinity? or what website(s) is the dataset made of?
Hi, for the furry part is mainly e621
Is there a plan to train danbooru artist style and copyright characters for this model?
Yes, that is the idea, but traing characters and artistas has been challenging so far. Focusing more on artists.