JanK2: Krea 2 Anime Checkpoint

JanK2 is a fine-tune of Krea 2 Turbo, trained to output anime style by default. No need to prompt for "anime," "anime style," or similar tags to get there.

The training set combines real data pulled from JANKU (my Illustrious/NoobAI checkpoint) with a small, hand-picked set of community-favorite images. The JANKU data anchors the anime aesthetic and linework consistency you'd expect from that line, while the tiny curated set was used to nudge composition and detail quality rather than to broadly retrain the model. This is a light, targeted fine-tune on top of Krea 2's own strong base capability, not a from-scratch retrain.

Because it inherits Krea 2's underlying strengths, you still get the wider style range, better environments/backgrounds, and stronger prompt adherence that Krea 2 is known forr, just steered toward anime as the default output instead of requiring style tags every time.

How to Get Started with the Model

You'll need a webui/frontend that supports Krea 2 and your chosen quantization:

• ComfyUI: v0.27.0 or later (v0.28.0+ if you're using an INT4 ConvRot build)

• Forge Neo: 2.27 or later

Grab the checkpoint file that matches your VRAM budget (INT8 ConvRot is a good default for most 8-16GB cards) and drop it into your usual checkpoints folder.

Both INT8 and INT4 is uploaded!

Safety / Censoring Notes

No uncensoring is baked into this checkpoint itself. If you want to adjust content restrictions, that's handled on your own end, through whatever LoRA or workflow node you'd normally use for that, rather than something this model applies automatically.

Worth calling out: whatever method you use to uncensor or loosen restrictions can sometimes cost you a bit of image quality (softer detail, occasional composition drift). It's usually better to leave that off by default and only switch it on for the specific generations where you actually need it, rather than running it on every prompt.

Recommended Settings

Follows Krea 2 Turbo's official defaults:

• Resolution: 1024-2048px

• Steps: 8

• CFG scale: 1

• Sampler: euler, er_sde

• Scheduler: simple, normal

Prompting Guidelines

Refer to the general Krea 2 prompting guidelines for structure and phrasing. Those still apply here.

I recommend checking out Gladas's workflow: https://civarchive.com/models/2807513/gladas-workflow-for-krea-2-turbo

Credits / Notes

• Base model: Krea 2 Turbo (Krea AI)

• Training data: Subset of JANKU training data + small community-favorite curated set

• License: https://github.com/krea-ai/krea-2/blob/main/LICENSE.md