Experimental SDXL anime checkpoint, trained with Direct Preference Optimization against an in-house preference set focused on body / anatomy fixes (fixbody-dpo-captioned, fixbody-dpo-danbooru).

Part of the "Niku" line, which traces back to NikuNiku900 — a long-running Schneewolf Labs anime-art project. NikuXL v0.1 was the first DPO experiment in the line, exploring whether preference learning could nudge an SDXL anime mix toward better anatomy and softer rendering without retraining the whole stack.

Notes

Outputs lean toward soft, painterly anime portraiture; reasonable hands and anatomy by SDXL standards.
Treat as an artifact rather than a reproducible release — the original training run isn't preserved in detail.
The training pipeline this came out of has since been rewritten and folded into Atelier, Schneewolf Labs' diffusion-training framework. Active work lives there.
Released because it served as the v1 image-gen backbone for flammen.ai, and a clean upstream pointer is more useful than a stale local checkpoint.

Usage

Standard SDXL checkpoint — works in any SDXL-compatible runner (AUTOMATIC1111, ComfyUI, diffusers).

python

from diffusers import StableDiffusionXLPipeline
import torch

pipe = StableDiffusionXLPipeline.from_single_file(
    "NikuXLv0.1.safetensors",
    torch_dtype=torch.float16,
).to("cuda")

image = pipe(
    "1girl, anime style, masterpiece, soft lighting, portrait",
    negative_prompt="lowres, bad anatomy, blurry, photorealistic, 3d",
    width=1024,
    height=1024,
    num_inference_steps=28,
    guidance_scale=7.0,
).images[0]

License

Inherits the SDXL license (CreativeML Open RAIL++-M).

schneewolflabs/NikuXL-v0.1

NikuXL v0.1

Notes

Usage

License

schneewolflabs/
NikuXL-v0.1