FLUX 1 D:FLUX provides better image quality and language understanding. Although named Sky Bar, it only continues the original intention and can actually generate multiple types of barsSDXLThe SDXL version enhances semantic understanding and can produce more interesting contentAlthough named skybar, it can actually generate more bar styles for everyone to tryCheck recommended jugernaut_XLora weight recommendation of 0.5-0.7If you need a high-altitude bar, you can refer to these words：city,city lights,cityscape,outdoors,rooftop,scenery,sky,skyscraper,cloudy sky,the bar on the rooftop,champagne,skyline,balcony,I copied the prompt words from the previous 1.5 in the first few reference images, For the other style reference images, I wrote the prompt words using GPT. Below is the reference template:skybar,&lt;lora:skybar:0.6&gt;, (wine cabinet:1.2),(bar counter:1.2),(wine bottle:1.2),(wine glass:1.1),- Features:- Atmosphere:- Style:- Background:- Lighting:SD1.5 model_2.0Added more outdoor bar material training,The generalization with other LORAs has been adjusted to produce more interesting effects with more LORAs,Realistic classes are recommended for large modelsThe picture training set is 768*512 so it is recommended not to exceed,If anyone can fix it with ADetailer,step 30-40 is recommended,lora strength is recommended at 0.6-0.8Drawing is recommended to open HD repair about 0.6

Sky Bar

sky bar:The 1.0 version of the model is easy to produce close shots, while the 1.01 version is easy to produce mid range shots. Same training data for both versions, so I uploaded both ，I personally feel that V1.0 has a better fusion of various prompt word，weights ranging from 0.5 to 1.0. Training use 768×512. The dataset is all images of unmanned scenes in bar, so the effect of a separate unmanned scene image is better. If you want to add man, you can reduce the weight. Adding characters such as Lora or adding characters/nightclubs/music festivals to the prompt words can also produce some special effects, Remember to use Adetailer to repair the face. The online image quality is not 4k, and if some images are unclear, high-definition repair can be added.1.0版本的模型容易出近景，1.01版本的容易出中景，训练参数是一样的，所以我两个都上传了，个人感觉V1.0对各种提示词的融合度更好一些，权重0.5-1.0都可以，训练使用768×512，数据集都是清吧无人的场景图片，所以单独无人场景图效果好一些，想添加人物就降低权重，加上人物Lora或者在提示词加上人物/夜店/音乐节之类的话也会出一些特殊效果，人脸记得要开用Adetailer来修复，网上图片画质不是4k，有的成图不清晰的话可加高清修复。

Files with SHA256: 3964da825804a3a1ce51e199893a7ea537cbb23a5ee398b32f525c86b236e057