CivArchive
    Preview 144726491
    Preview 144726492
    Preview 144726493
    Preview 144726497
    Preview 144726499
    Preview 144726500
    Preview 144726501
    Preview 144726502
    Preview 144726503
    Preview 144726504
    Preview 144726505
    Preview 144726506

    Anime / illustration → photorealistic transform

    Edit LoRAs that turn anime-style or illustrated images into realistic photographic style while following the input composition. One version per base model — currently MOSS-Image, Krea 2 and Qwen-Image 2.1; start from the newest.

    • Prompt guidance is effective across versions: the simplest transform to realism / 转成写实照片 works, and extra prompt words can steer lighting, mood, clothing color and material — even character age — within a range.
    • Every preview image embeds its ComfyUI workflow — drop one onto the canvas to load it.
    • Sampling settings matter: sampler, scheduler, denoise (it does not have to be 1) and latent size all influence the result. Experiment freely.

    Non-commercial use only. Re-uploads and ad-farm mirrors are not tolerated.


    动漫 / 插画 → 写实照片转换

    将二次元或插画图像转换为写实照片风格的编辑 LoRA,遵循输入构图。每个底模一个版本——目前有 MOSS-Image、Krea 2、Qwen-Image 2.1,建议从最新版本用起。

    • 各版本对提示词引导均有响应:最简单的 transform to realism / 转成写实照片 即可生效;追加提示词可以在一定范围内调整光影氛围、服装颜色材质,甚至角色年龄。
    • 每张预览图都内嵌完整 ComfyUI 工作流,拖入画布即可加载。
    • 采样设置会影响结果:采样器、调度器、降噪率(不必是 1)、latent 尺寸都会带来变化,欢迎多尝试。

    本系列模型严禁商用(包括机翻打广告的搬运)。欢迎中文留言交流。

    Description

    EN — Preview release of the realism-transform edit LoRA trained on Qwen-Image 2.1.

    • Recommended LoRA strength: 0.8 – 1.2.
    • Keep the generation resolution at ≤ 2 MP total — higher totals cause deformation.
    • Strengths: native editing behavior with high controllability; background replacement and quality shifts respond strongly to prompts (especially with cfg > 1); no NSFW hallucination — it will not add NSFW content to clean images.
    • Known issues: (1) Qwen's limb weakness can make limbs unstable; (2) NSFW output needs help from other LoRAs, otherwise performance is poor; (3) eyes / expressions are best steered by prompts; (4) precision is below the Krea 2 version and the model is smaller — details drift if not controlled by prompt.

    中文 —基于 Qwen-Image 2.1 训练的转写实编辑 LoRA,预览版。

    • 推荐 LoRA 强度:0.8 – 1.2。
    • 生成分辨率总像素不超过 2MP,再高会出现变形问题。
    • 优点:原生编辑功能、可控性高;背景替换、画质变化对提示词响应敏感(cfg 大于 1 时尤甚);没有 NSFW 幻觉,不会在非 NSFW 图上凭空出现 NSFW 内容。
    • 已知问题:(1)Qwen 本身的肢体缺点会导致出图肢体不稳定;(2)NSFW 内容需要借助其他 LoRA,否则表现不佳;(3)人物眼神表情最好用提示词控制;(4)画面精度不如 Krea 2 版本、参数量较低,部分细节无提示词控制时会失控。

    FAQ

    Comments (3)

    jiaolongOct 7, 2026
    CivitAI

    我目前测试过的这么多的lora里看着最不违和,各方面一致性保持的最好的lora,除了脸上的纹身 ,不过不影响这是一个非常好的转真人lora,审美符合亚洲人

    2856313326146Oct 12, 2026
    CivitAI

    怎么下载你的工作流?我用我自己的工作流生成的图片很抽象,脸部崩坏

    2856313326146Oct 12, 2026
    CivitAI

    我已提取到了你的工作流,谢谢

    LORA
    Qwen 2.1

    Details

    Downloads
    1,785
    Platform
    CivitAI
    Platform Status
    Available
    Created
    10/5/2026
    Updated
    11/21/2026
    Deleted
    -
    Trigger Words:
    transform into realistic style

    Files

    realism_transform_qwen21_preview.safetensors