SDXL1.0开源了,相比以往的模型对文本的理解更加透彻智能,且图像质量更高更好,
下面我们将测试SDXL1.0对文本的理解这一能力。
测试方法:用Midjourney的prompt,并使用SDXL1.0模型直出图片,不使用lora,并对比Midjourney的图像效果。
具体方法:
这里我们使用在线的平台(吐司art:https://tensor.art/)来进行出图。
底模选择:SDXL 1.0-Base
反向提示词:bad_prompt_version2
采样方式:DPM++2M Karras
提示词相关性:7
其他参数默认
下面所有的提示词均来自Midjourney Discord服务器,选择了不同类型的图片提示词来进行出图对比(顶部横向图片为SDXL1.0出图,底部四宫格为Midjourney出图):
1.anime guy, short silver hair, black white silver gold clothes, comet, asteroids in background, comet in background, silver eyes, serious/neutral expression, serious/neutral face, forehead armor headband silver and gold, flying, cool, creative, powerful look, 4k
2.Wonderful and delicious breakfast on the terrace of a trendy restaurant. Pastel-colored flower bouquets. The morning sun.
3.anime - style woman in orange bodysuit laying on a bed, stunning sci-fi concept art, trending on juxtapoz magazine, cushart, technology flight suit, overwatch skin, insanely inflated hips, inspired by INO, james jean aesthetic, zerochan, wears tiny spacesuit v4
4.cat with kitten, disney cartoon style, white background
5.ancient griffin flying, body is surronded and filled with blue lightning mixed with gray, in a storm with blue lightning, and glowing illuminated blue eyes, flying sideways, cyberpunk, zoomed out a tiny bit, high quality
6.a female heavy cybernetic armor, blending its arm to a laser, bald, muscles, full helmet, glowing purple eyesight on helmet, purple laser spot, armored, dark red armor, chrome reflexion, neon green lining, style digital scenic, background black desert city, in motion, side view, full body, night, ultradetailed, ultrarealism, realistic photography 4k
7.soccer player in white jersey celebrating his goal, in the style of zeiss batis 18mm f/2.8
通过以上对比可以看出,SDXL1.0对语义的理解已经有了非常大的进步,且出图稳定了很多,二者出图的内容已经非常接近了,只是Midjourney在细节上依然更加出色,这些SDXL1.0可以通过Vae、Lora、Controlnet等来实现。
相比Midjourney的付费计划,SDXL1.0开源免费显然更香!