Full fine-tune of Krea 2 Raw (12.8B DiT) — not a merge, not a LoRA. Anthro and furry are its home turf, with real anime / kemono range, multi-character scenes that stay separate characters, and 1,113 artist styles you can call by name.

Current version is v0.5: 13 epochs at 512 px over a curated corpus, then an artist-focused pass, then a final pass at 1024 px. If you used v0.1 — the @artist tokens actually work now, and it trained a lot longer.

Settings

Steps 8
CFG 1.0
Sampler / scheduler euler / simple
Shift 1.15 (the ComfyUI default for Krea 2, no extra node needed)
Resolution 1024 native
CLIPLoader type krea2

CFG above 1.0 burns the image on the Turbo files, and at CFG 1.0 the negative prompt does nothing — don't bother writing one.

Prompting — read this bit, it isn't Illustrious

Training captions start with a rating tag, then the artist token, then the description. Prompt it the same way and you get the most out of it:

sfw, @artistname, A [medium] of [subject, species, appearance]. [Pose, action, expression]. [Clothing / body details]. The background features [setting, lighting, atmosphere]. [Style / rendering notes].

Ready to paste (swap the artist token for one from the list, or drop it entirely):

sfw, A digital painting of a grey wolf anthro woman standing on a rain-slick rooftop at night, one hand on the railing, looking back over her shoulder. She wears an oversized bomber jacket over a cropped top, tail curled against her leg. The background features neon signage blurred by drizzle and wet concrete reflecting magenta and cyan light, seen from a low angle. Soft cel-shaded rendering, crisp linework, painterly highlights.

Booru / e621 tag lists work too — I trained with tags this time:

sfw, @artistname, 1girl, kemonomimi, fox ears, white hair, blue eyes, kimono, holding umbrella, snow, night, detailed background, looking at viewer, upper body

• Rating tag first: sfw or nsfw. Artist token right after it.

• 60–120 words is the sweet spot for prose. Short prompts work, detail works better.

• You don't need SDXL-style quality boosters here. Describing the image does more.

• Ask for the background. If you don't describe it, the model fills it in and you can get a flat or blurred one.

• One- or two-word prompts sample the model's whole range, so expect a random style.

Artist styles

1,113 tokens. Exact match, @ included, placed right after the rating tag. Full list: ARTISTS.md on Hugging Face.

Artists with more images in the set respond harder. A token also pulls in the kind of subject that artist draws, not only the rendering — so state your subject explicitly if you want something else.

Which file do I download

Civitai renames the files on download, so go by size:

• fp8, 11.9 GB — ComfyUI. The one most people want.

• GGUF Q8, 12.8 GB — needs the molbal/ComfyUI-GGUF fork; city96's node doesn't support Krea 2 yet.

• int8 ConvRot, 12.6 GB — forge-neo and other int8 runtimes.

• int8, 12.6 GB — plain int8, same size, no ConvRot.

• W4A8 ConvRot, 7.2 GB — the smallest one. Needs ComfyUI 0.31+ (native W4A8 loader), an SM 8.0+ GPU and PyTorch cu130+.

• base 0.5v, 23.9 GB bf16 — the Non-Turbo checkpoint. This is the one for LoRA training and further fine-tuning, or to pick your own Turbo LoRA strength. Without the LoRA it wants 52 steps at CFG 3.5.

bf16 Turbo, Non-Turbo int8 and SHA256 checksums for every file are on Hugging Face.

ComfyUI setup

• The fp8 file goes in ComfyUI/models/diffusion_models/

• Text encoder qwen3vl_4b_bf16.safetensors from Comfy-Org/Krea-2 goes in models/text_encoders/

• VAE qwen_image_vae.safetensors from the same repo goes in models/vae/

• Drag the attached workflow JSON into ComfyUI and hit queue.

No GPU?

vaelico.ai runs v0.5 in the browser, free tier included.

Known issue: stretched top / bottom edge

Some generations come out with the top or bottom edge smeared or stretched. It's my bug, not yours — not a ComfyUI or VAE decoding problem, so don't go hunting for it in your setup.

Cause: I chose not to auto-crop anything during training, so images that didn't match a training aspect ratio were letterboxed, and the letterbox was filled by replicating the edge pixels. That fill was masked out of the loss, but the latent row sitting on the boundary still carried a fraction of it — so the model picked up "edges sometimes stretch". With the bucket set I used, only about a quarter of the images filled the canvas exactly, which is why it shows up as often as it does.

It's fixed in the data pipeline for v1 (many more aspect buckets, so the vast majority of images fill the canvas with real pixels, plus a small centred crop for the rest). It can't be patched into v0.5 — these weights already learned it. For now: crop the affected rows or reroll the seed.

Other rough edges, and what's next

• Multi-character holds up far better than the generalist models, but busy compositions still miss a fair share of the time.

• Backgrounds need asking for, as above.

• LoRAs trained on base Krea 2, or on Wulver v0.1, may behave differently here — the weights have moved.

v1 is in the works: more data, and the pass this needs on structure and composition. It'll show up on this page when it's ready.

Credits and license

Base model and the Turbo LoRA: Krea. GGUF conversion: molbal's ComfyUI-GGUF fork.

This is a modified version of the Krea 2 model by Krea. Krea 2 is licensed under the Krea 2 Community License Agreement — by using this model you agree to its terms, including the Acceptable Use Policy. Not affiliated with or endorsed by Krea.

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.
by Vaelico