Stable Diffusion WebUI 核心文件夹解析与模型推荐

• BLIP：'看图说话'专家。用于反推提示词功能，分析图片内容生成 Prompt。
• generative-models：SDXL 模型架构支持。让 WebUI 能够运行新一代 SDXL 大模型。
• k-diffusion：采样器仓库。决定噪点如何一步步变成清晰图像，如 Euler a, DPM++ 2M Karras 等算法均在此。
• stable-diffusion-stability-ai：老版本核心架构。驱动 SD 1.5 / 2.0 等经典模型。
• stable-diffusion-webui-assets：静态资源与字体文件，保证界面在无网环境下也能正常显示文字。

模型推荐

Nova Anime XL【二次元】

链接：Nova Anime XL - IL v15.0 | Illustrious Checkpoint | Civitai 类型：Checkpoint (大模型/底模)。它不是 Lora，不需要挂载在其他模型上，直接选择即可。 架构：SDXL。注意这不是老旧的 SD 1.5 模型。 要求：模型较大（约 6GB），推荐分辨率 1024x1024。显存低于 8GB 的运行起来可能会比较吃力。

配置示例

正向提示词：

text
masterpiece, best quality, amazing quality, 4k, very aesthetic, high resolution, ultra-detailed, absurdres, newest, scenery, colorful, saturated, anime, anime coloring, 1girl, solo, cute, pink hair, long hair, choppy bangs, long sidelocks, nebulae cosmic purple eyes, rimlit eyes, facing to the side, looking at viewer, downturned eyes, light smile, red annular solar eclipse halo, red choker, detailed purple serafuku, big red neckerchief, (glowing stars in hand:1.2), fingers, dispersion \(optics\), from side, from below, dutch angle, portrait, upper body, head tilt, colorful, rim light, backlit, (colorful light particles:1.2), cosmic sky, aurora, chaos, perfect night, fantasy background, dreamlike atmosphere, BREAK, detailed background, blurry foreground, bokeh, depth of field, volumetric lighting

负向提示词：

text
photorealistic, realistic, 3d, multiple views, multiple angle, split view, grid view, two shot, outside border, picture frame, framed, border, letterboxed, pillarboxed, 2koma, modern, recent, old, oldest, cartoon, graphic, text, painting, crayon, graphite, abstract, glitch, deformed, mutated, ugly, disfigured, long body, lowres, bad anatomy, bad hands, missing fingers, extra fingers, extra digits, fewer digits, cropped, very displeasing, (worst quality, bad quality:1.2), sketch, jpeg artifacts, signature, watermark, username, (censored, bar_censor, mosaic_censor:1.2), simple background, conjoined, bad ai-generated

关键参数：

• Steps: 20
• Sampler: Euler a
• CFG scale: 4.5
• Size: 768x1344 (可根据需求调整至 1024x1024)
• Clip skip: 2
• Hires steps: 20, Upscale: 1.5

实际运行时，SDXL 模型对显存有一定要求，如果显存不足，可以尝试降低分辨率或使用优化插件。