Vaelico/Wulver

<div align="center"> <img src="assets/vaelico-logo.png" width="72" alt="Vaelico">

# Krea 2 - Wulver _v0.5

Full fine-tune of Krea 2 Raw (12.8B DiT) — anthro & furry, with strong anime / kemono range,<br> true multi-character composition and 1,113 artist styles you can call by name.

Civitai page · vaelico.ai </div>

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.

What's new in v0.5

• —Much longer training. 13 full epochs over the curated corpus at 512 px (384,672 samples per epoch, about 156k steps at batch 32), the 13th with the learning rate annealed to zero. Then an artist-focused pass at 512 px and a final curated pass at 1024 px.
• —Artist tokens work. 1,113 @artist tokens, listed in ARTISTS.md. In v0.1 the @ prefix was ignored.
• —Turbo from the official Krea 2 Turbo LoRA, merged into the Turbo files at strength 1.0 (8 steps, CFG 1). The Non-Turbo files ship too, so you can pick your own strength or skip it.

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 — see prompting tips below

Files

GGUF quants are on the Civitai page.

Settings (Turbo files)

Other recipes (Non-Turbo file)

The Turbo LoRA is krea2_turbo_lora_rank_64_bf16.safetensors from Comfy-Org/Krea-2 (loras/), loaded with a LoraLoaderModelOnly node.

Sampler euler / simple in all three. Without the LoRA (CFG above 1) use an empty CLIPTextEncode as the negative, not ConditioningZeroOut.

ComfyUI setup

1. 1.Wulver_v0.5_fp8_e4m3fn.safetensors → ComfyUI/models/diffusion_models/
2. 2.Text encoder qwen3vl_4b_bf16.safetensors from Comfy-Org/Krea-2 → models/text_encoders/
3. 3.VAE qwen_image_vae.safetensors from the same repo → models/vae/
4. 4.Drag Wulver_workflow.json into ComfyUI and hit queue.

Prompting tips

Training captions start with a rating tag, then the artist token if there is one, then the description. Prompting the same way gets you the most out of it:

sfw, @artistname, A digital painting of [subject & species, appearance details]. [What they are doing, pose, expression]. [Clothing / body details]. The background features [setting, lighting, atmosphere].

• —Detailed descriptive prose works best; booru-style tag lists also work
• —Artist tokens must match ARTISTS.md exactly, @ included. Put them right after the rating tag
• —One- or two-word prompts sample the model's whole range, so expect random styles. Describe what you want
• —At CFG 1.0 the negative prompt has no effect

Compatibility notes

• —Same architecture as Krea 2 Raw: Qwen3-VL-4B text encoder, Qwen Image VAE, stock ComfyUI loaders
• —w4a8-convrot needs ComfyUI ≥ 0.31.0 (native W4A8 loader), an SM 8.0+ GPU and PyTorch cu130+
• —LoRAs trained on base Krea 2 or on Wulver v0.1 may behave differently on v0.5 (weights have moved)

Checksums and provenance

All v0.5 files, including the Civitai-only GGUF (Q8CR / Q50 / Q40) and plain int8 files, with sizes: [SHA256SUMSv0.5.txt](SHA256SUMSv0.5.txt). The v0.5 Non-Turbo checkpoint and how each file was made: [NONTURBOv0.5.md](NONTURBO_v0.5.md).

Previous version

The v0.1 files (Wulver_v0.1_*) stay in this repository unchanged. See NON_TURBO.md for the v0.1 Non-Turbo checkpoint.

License

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). The full license text is included as LICENSE.md and the attribution notice as NOTICE in this repository. Not affiliated with or endorsed by Krea.

<div align="center"> <img src="assets/vaelico-logo.png" width="40" alt="Vaelico"><br> <em>by Vaelico — generate online shortly at <a href="https://vaelico.ai">vaelico.ai</a></em> </div>