Nick Wilde <尼克王尔德> character - zootopia - krea2 raw

ENGLISH / 英文说明

Nick Wilde character LoRA for Krea 2 RAW BF16, by RiverL Animation.

Trigger word: Nick_Wilde

Base model: krea2_raw_bf16.safetensors

LoRA rank: 32 | Saved precision: BF16

RECOMMENDED USE

Best suited to head portraits and upper-body images, including frontal, profile, and three-quarter views. Outfit changes generally worked well in the author's tests.

Start with a LoRA strength of 0.8 and a resolution of 1.0 or 1.7 megapixels. Four checkpoints are available: 1500, 1800, 2100, and 2400 training steps. Load one checkpoint at a time. These numbers refer to training steps, not image-generation sampling steps.

KNOWN LIMITATIONS

Full-body results may have distorted proportions or incorrect hands because the training data contains few complete full-body examples. If you want to try, use a strength of 0.8. Avoid prominent views of palms, soles, or feet.

Many training images are heavily cropped, so parts of the character may appear outside the frame. Complete framing is not guaranteed.

Happy expressions can be unreliable. Try the 1800- or 2100-step checkpoints first.

The training images have limited sharpness. Avoid excessively high resolutions. Do not put "21:9", "Ultra Wide", or "ultrawide" in the prompt, as these may produce black bars. Set the aspect ratio in the workflow's resolution node instead.

The LoRA was trained with BF16. Results with quantized or pruned base models are not guaranteed. The author observed approximately 30,000 MB of VRAM usage on an RTX 5090 with the BF16 workflow; actual usage depends on the setup.

Please read the included README and annotated ComfyUI workflow for setup instructions. Recommendations are based on the author's tests; results vary with prompts, seeds, and settings.

USAGE TERMS

Personal, non-commercial use does not require attribution.

If you publish content generated using this LoRA online, you must either link to this model's original Civitai release page, link to the author's profile below, or include the credit "Civitai@RiverL Animation".

Commercial use of any kind is prohibited.

Do not redistribute, reupload, or share the LoRA files. Direct other users to this original release page on Civitai. Sharing the original release page link is allowed.

Author profile: https://civitai.com/user/RiverL_Animation

中文说明 / CHINESE

适用于 Krea 2 RAW BF16 的《疯狂动物城》尼克·王尔德角色 LoRA。

作者：RiverL Animation https://space.bilibili.com/3461563421821040

触发词：Nick_Wilde

底模：krea2_raw_bf16.safetensors

LoRA rank：32｜保存精度：BF16

使用建议

────────────────────────────────────────

最推荐生成头部肖像和上半身图片，包括正面、侧面和四分之三侧面。在作者的测试中，服装替换总体表现良好。

建议先使用 0.8 的 LoRA 强度，分辨率设为 100 万或 170 万像素。目前提供 1500、1800、2100 和 2400 训练步数的四个版本，每次加载其中一个即可。文件名中的步数是训练步数，与生成图片时的采样步数不同。

已知问题

────────────────────────────────────────

训练数据中的完整全身图片较少，因此全身生成可能出现身体比例不协调或手部错误。如果仍想尝试，建议使用 0.8 的强度。尽量避免突出展示手掌、脚掌或脚部。

许多训练图片存在严重裁切，因此生成结果可能有部分角色位于画面之外，不能保证完整出镜。

开心表情的表现不太稳定，建议优先尝试 1800 步或 2100 步版本。

训练图片清晰度有限，不建议使用过高的分辨率。不要在提示词中写“21:9”“Ultra Wide”或“ultrawide”，这些词可能导致图片上下出现黑边。请通过工作流中的分辨率节点设置画幅比例。

本 LoRA 使用 BF16 底模训练，不保证搭配量化或剪枝版底模时的生成效果。作者在 RTX 5090 上运行 BF16 工作流时，观察到约 30,000 MB 的显存占用；实际占用会随运行环境和设置变化。

具体安装和操作方法请参阅附带的 README 与带注释的 ComfyUI 工作流。以上建议来自作者的测试，实际效果会随提示词、种子和设置变化。

使用条款

────────────────────────────────────────

个人非商业使用不要求署名。

如果将使用此 LoRA 生成的内容发布到网上，必须任选以下一种方式署名：提供本模型在 Civitai 的原始发布页链接、提供下方作者个人主页链接，或注明“Civitai@RiverL Animation”。

禁止任何形式的商业使用。

不得转发、重新上传或分享 LoRA 文件。如有其他人需要使用，请引导其到本模型在 Civitai 的原始发布页下载。可以分享原始发布页链接。

作者个人主页：https://civitai.com/user/RiverL_Animation