CivArchive
    【KREA2 Hybrid Style】Exquisite Sculpted · Delicate Miniature Figure Style 【KREA2混合风格】玲珑凝塑·精致小巧手办风 - v1.0
    NSFW
    Preview 144831685
    Preview 144831693
    Preview 144831749
    Preview 144831744
    Preview 144831782
    Preview 144831785
    Preview 144831821
    Preview 144831826
    Preview 144831857
    Preview 144831874

    本 LoRA 并非单一的文生图微调,而是由多个风格 LoRA 融合,凝练提取出的复合风格lora。

    核心美学在于“玲珑凝塑”的极致立体感与“精致小巧”的无瑕造型。加载此模型后,画面彻底摒弃了现实中的杂乱与随机光影,人物肌肤与服饰轮廓被精细雕琢,呈现出如同高级潮玩般完美的三维形态。线条圆润流畅,带有纯粹且均匀的视觉铺陈,环境背景干净通透,常伴随全息柔和的底色,营造出一种超脱现实、如同精致模型般静谧且极具可塑性的视觉氛围。

    This LoRA is not a single text-to-image finetune, but a composite style LoRA refined from a fusion of multiple style LoRAs.

    Its core aesthetic lies in the ultimate three-dimensionality of “exquisite sculpting” and the flawless form of “delicate miniature.” When loaded, the image completely abandons the clutter and random lighting of reality; character skin and clothing contours are finely sculpted, presenting a perfect three-dimensional form akin to high-end art toys. The lines are rounded and smooth, with a pure and even visual presentation. The environmental background is clean and translucent, often accompanied by holographic soft base tones, creating a surreal, serene, and highly malleable visual atmosphere reminiscent of a finely crafted model.

    Description

    FAQ

    Comments (20)

    youtougle2422Oct 6, 2026
    CivitAI

    Looks great...but how is it triggered?

    There's no trigger words shown, there's no prompts on the images, there's no embedded metadata in the images... how do we guide its usage? 🤔

    baizuobusi
    Author
    Oct 6, 2026· 1 reaction

    Glad you like it! No trigger words needed by design. The LoRA provides the style on its own, so just describe what you want naturally.

    UnicomOct 6, 2026
    CivitAI

    Is it LoRA? Dont working on my setup (blue screen).

    baizuobusi
    Author
    Oct 6, 2026

    Hello. I have already responded to your similar comment on my other model, but you've left it again here. I strongly suspect this is a configuration issue on your end. Could you please clarify your base model and the settings you're using?

    UnicomOct 6, 2026

    @baizuobusi 
    I use the DrawThings app on macOS. I never have issues with LoRAs. There are sometimes problems with LyCORIS models—specifically, they might fail to import—but your model does import into DrawThings; however, its file size doesn't match that of other LoRAs of the same rank. If it is indeed a LoRA and not a DoRA or LyCORIS, this is the first time I’ve encountered such a glitch—and it happens only with your models (in over three years, I’ve used thousands of LoRAs across all kinds of architectures—SD1.5, SDXL, FLUX, ZIT, Klein, Krea2, Qwen—and never had issues like this). I’d like to figure out the cause.

    baizuobusi
    Author
    Oct 7, 2026

    @Unicom First, this is a Krea2-based LoRA created by merging multiple LoRAs using weight difference subtraction. It maintains a rank of 32, and the file size is consistent with my other directly trained 32-rank Krea2 LoRAs.

    Second, I do not use macOS. I have tested this model extensively in ComfyUI on an NVIDIA environment with zero issues. Since I cannot reproduce this problem and do not have access to a macOS/DrawThings setup, I am unable to assist further with this specific issue. It appears to be a compatibility limitation with your specific environment.

    UnicomOct 7, 2026

    @baizuobusi Are you saying that you merged the LoRAs into the checkpoint and then extracted a LoRA using SVD? If so, did you use a BF16 checkpoint or a quantized one?

    baizuobusi
    Author
    Oct 7, 2026

    @Unicom Yes, you are correct about the SVD extraction process. To clarify exactly how I extracted this: I did not save a merged checkpoint first. Instead, I calculated the difference directly in memory between the 'base model' and the 'base model + LoRA chain' using ComfyUI nodes, and the entire extraction process (both the difference calculation and SVD decomposition) was performed strictly in FP32 precision. Even though the base model was loaded in INT8, the code upcasts it to FP32 before subtraction. This method avoids any intermediate precision loss that would occur from exporting and reloading a merged checkpoint , meaning the extraction is mathematically precise and should not be the cause of your issue.

    I have tested this extensively in ComfyUI on an NVIDIA environment with zero issues. Since I do not use macOS, I cannot verify its behavior on your specific setup, and therefore I cannot provide any advice or suggestions regarding Mac or DrawThings. It appears to be a compatibility limitation with your specific environment.

    UnicomOct 7, 2026

    @baizuobusi It is almost certain that the issue stems from the fact that the base model was quantized and then dequantized. I have worked extensively with SVD and differential LoRAs; frequently, when a quantized base model was used (especially if it wasn't FP8), the dequantization process failed, resulting in either a "blue screen" or a mosaic of noise during generation. I suspect you used "INT8 convrot." To dequantize from that format, you need to know exactly how the quantization was performed; I am unable to dequantize that format on a Mac. It is likely that your LoRA works for you because the quantized base model also worked in your specific setup, but it fails on other configurations (such as Macs or systems using a different GPU manufacturer). If possible, try avoiding quantized checkpoints and perform the merge using BF16—or, as a last resort, use FP8 base quantization (though there are no guarantees there). This approach should yield a universal LoRA.

    baizuobusi
    Author
    Oct 7, 2026

    @Unicom You might be right about the dequantization issue. But I don't use BF16, and I have no obligation to adapt my workflow for an environment I don't use. It works perfectly in my setup.

    UnicomOct 7, 2026

    @baizuobusi This means your LoRAs will work only for you and for those with a setup similar to yours. Furthermore, merging with a quantized base model introduces significant quality degradation. Proper dequantization prior to merging could at least prevent errors from accumulating exponentially, but that is not the case here.

    baizuobusi
    Author
    Oct 8, 2026

    @Unicom I have no obligation—and it's not worth my time—to change my workflow for this. If my models are incompatible with your setup, simply don't use them. I don't need a lecture.

    hornysilverhandOct 7, 2026
    CivitAI

    It is creators like you who make the community so vibrant. I’ve tried out a few of your LoRAs; they all perform well without any loss of detail, though I did notice varying levels of sensitivity to prompts—such as when generating "full-body" shots. Also, the facial resemblance is very strong, giving the characters a slightly youthful look, but the style remains excellent. Great work.

    baizuobusi
    Author
    Oct 7, 2026

    Thanks! Appreciate the feedback. Glad you're enjoying them.

    TianNai713Oct 8, 2026
    CivitAI

    大佬什么时候能做个半写实的风格就好了

    baizuobusi
    Author
    Oct 8, 2026

    正愁选题呢,详细说说?

    TianNai713Oct 12, 2026

    半写实CG风格那种。感觉如果能做个那种类似漫转真的更好,目前漫转真的挺多但是转成半写实的貌似没有......

    baizuobusi
    Author
    Oct 12, 2026

    @TianNai713 我最新的那个模就是偏cg风的,漫转真用我的那个S2A也能做,有专门做真实系的real变体,加风格lora就是风格转绘

    fzcxhyjOct 12, 2026
    CivitAI

    很喜欢,不过生成的图是非商业用途的吗?

    baizuobusi
    Author
    Oct 12, 2026

    不好意思,非商用哈

    LORA
    Krea 2

    Details

    Downloads
    213
    Platform
    CivitAI
    Platform Status
    Available
    Created
    10/6/2026
    Updated
    11/20/2026
    Deleted
    -

    Files

    【KREA2混合风格】玲珑凝塑·精致小巧手办风.safetensors