琢磨了一天,不知道有没有人发过.
如题:
正常生成太慢,而使用turbo又会导致提示词理解变差
可以先正常生成一张低分辨率的图片,再使用加载了加速lora的模型进行放大,这样就可以保持原生的提示词理解了
个人调试的参数如下:
可以看到,使用加速lora加速后,只需要4步
模型:silvermoonmix,这个模型的肢体崩坏比较少一些
加速lora推荐3个,不推荐官方推荐的那个,会让画面变得很乱:
2619830
2466415
2364703
分别进行了二次放大:
分辨率是768x960
放大1.33倍1024x1280
放大1.25倍1280x1600
生成时间对比:
4060laptop
此方法768x960,25步 | 44秒 |→ 1024x1280,4步 | 57秒 |→ 1280x1600,4步 | 1分14秒|
直出1024x1280,30步 | 1分39秒
直出1280x1600,30步 | 2分22秒

为什么我是10秒

楼主用的哪个加速lora？第一个404了

怎么做到头发画这么细又不粘连的

等wai版出来测试一波

如果二采会掉画师串的话，应该适当增加二采的cfg。

楼主我二采出来细节会糊 我是单次1.5倍放大 0.65降噪 还是xl的习惯 是参数的问题吗？

我个人感觉还是官方的0.2比较好（画面乱确实是个缺点），这几个lora我都用过，但是不知道为什么它们对无lora画师的画风基本不怎么遵循，更加偏向特定画风，除此之外其他地方我没怎么测试过，我用的是0.2官lora0.7强度，加了个美学lora，Aesthetic Quality Modifiers强度1.0，

我和你一样配置，没加速LoRA时间和你加速差不多，用了15秒左右1024x1024，35步直出60 70秒左右

来支持了

感觉很好用，谢谢
lz

测试精度最好跑一些多人大场景，脸部修复能力和细节黏连修复能力一目了然，跑头像不太明显，就算sd1.5跑头像也不至于脸部崩坏。

测试参数和提示词masterpiece, best quality, score_9, score_8, score_7, year 2025, newest, highres, absurdres, very aesthetic,
on mountain, in forest, beautiful detailed water, realistic nature view, cinematic lighting, reflection light,
detailed background, photo background, dutch angle, feet out of frame,
(@namori:2),
yuru yuri, 4girls, joyful, walking together in a tight group on a mountain trail,
(The first girl, (akaza akari is wearing a white ruffled blouse with lace trim at the collar and a dark navy pleated skirt:1.6). She has short red hair tied in two round buns with small red ribbons. She is in the immediate foreground. Her shoulder and the back of her head fill the right edge of the frame. She is turning her head to look back:1.1),
(The second girl, (toshinou kyouko is wearing an orange and yellow striped sweater with a loose cowl neck and distressed denim shorts:1.6). She has blonde hair with a colorful hair clip on one side. She is walking just behind. One hand is reaching forward. Her mouth is open as she talks:1.1),
(The third girl, (funami yui is wearing a dark blue sailor-style blouse with a white collar and a light grey pleated skirt:1.6). She has short black hair with a thin white headband. She is walking with her hands in her pockets. A slight smile is on her face:1.1),
(The fourth girl, (yoshikawa chinatsu is wearing a pink hooded jacket with white drawstrings over a cream-colored tank top:1.6). She has short pink hair with a small ahoge and two tiny hair bobbles. She is walking at the back. A small smile is on her face:1.1),
pine trees, dappled sunlight, depth of field, foreground background size contrast, wide angle perspective,

返图支持一波（表情不一样是因为过了只有质量词的面部细化，整体保留了一采的多数信息）