听取了zzz哥的建议,给模型起了个易于搜索辨识的名字,模型内容与先前版本相同。
本模型可画出多个景观上的概念,希望大家可以玩的开心。
prompt:
1.花卉:garden,flower搭配其他都可, 搭配gravel road可画出灰色砂石路
2.驿站:yz_station, service station,搭配其他建筑类相关词汇,可出驿站相关
3.古典园林:SZYL,east asian architecture, architecture,可出古典园林,搭配其他词汇出相应风格组团,石头等
4.秋景:autumn, autumn leaf tree, 搭配其他词可出现秋色叶树种与色调
5.夜景灯光:night,light,搭配其他词可出现深夜夜景灯光效果
6.雨景:rain,fog,多输入几次加大权重,搭配其他词汇可出现雨景
7.随机条件下可输入:ZHIGAN,作为触发词,提高画面的效果图风格
8.在材质上,对木头(wooden)、砖块(brick)有训练
建议使用长句描述自然语言,出不来的东西多输入几次
建议宽图,1024*576为好,cfg7及以下,如果加入lora后过拟严重可试试将clip跳过层调到4.6
可以选择开启动态cfg系数(一个插件)来降低画面的过拟合
可以搭配bad-picture-chill-75v,等负面embedding使用
可以用gpt自动输入的prompt,效果算是良好
搭配controlnet使用效果也算令人满意,作为国内的学生作业我认为是十分够用的
作为一个自主炼制并混合的景观大模型,本模型有以下特点:
本模型是一个定向性质的有些许过拟合的模型(几个词就能输出一张好看的图片),并不够十分泛化,画风也常常过拟,从大模型的高标准评判中来说,并不属优质,但想了想还是发出来,让更多的人看到景观也能用sd画出好图,给予更多的人一些思考。
本模型适合以下人群出图:刚接触sd想立即上手出好图的人、不会自主训练lora的人、想练习prompt使用与抽好看的卡的人、想用一个大模型抓紧时间画出几张好图的人。
本模型不适合以下人群出图:在摸索过许多模型后想自己训练lora模型的人、在度过新手期后想通过自己的训练来长期搭配大模型出图的人(这里我仍旧推荐chill)、下载了一个接近过拟合的模型以后就以为sd的效果主要依靠大模型决定的人。
本模型会用在之后的一些国内课程中。
作者qq名:冰醋酸(如果有群友认识我,嘿嘿)
I listened to the suggestion of Brother ZZ and gave the model a name that is easy to search and identify. The content of the model is the same as the previous version.
This model can draw multiple landscape concepts, and I hope everyone can have a good time playing.
Prompt:
1. Flowers: Garden, flower can be paired with anything else, and gray gravel roads can be drawn when paired with level roads
2. Post station: yz_ Station, service station, combined with other architectural related vocabulary, can be used for post station related
3. Classical gardens: SZYL, east Asian architecture, architecture, can be used to create classical gardens, combined with other vocabulary to create corresponding style groups, stones, etc
4. Autumn scenery: autumn, autumn leaf tree, paired with other words to create autumn leaf trees and tones
5. Night lighting: night, light, combined with other words to create a late night lighting effect
6. Rain scene: rain, fog, input multiple times to increase weight, and use other words to create a rain scene
7. Under random conditions, you can input: ZHIGAN as a trigger word to enhance the visual effect style of the screen
8. In terms of materials, there is training in wood and brick
It is recommended to use long sentences to describe natural language, and input more times for things that cannot come out
Suggest a wide image, 1024 * 576 is better, with cfg7 and below. If there is severe overfitting after adding Lora, try adjusting the clip skip layer to 4.6
You can choose to enable dynamic cfg coefficient (a plugin) to reduce overfitting of the image
Can be used with bad-picture-chill-75v or other negative embeddings
The prompt that can be automatically input using GPT has a good effect
The effect of using it with Controlnet is also satisfactory, and as a domestic student assignment, I think it is very sufficient
As an independently refined and mixed landscape model, this model has the following characteristics:
This model is a somewhat overfitting model with directional properties (just a few words can output a beautiful image), and it is not very generalized enough. The painting style is often overfitting, and from the high standard evaluation of large models, it is not of high quality. However, after thinking about it, it is still published to allow more people to see the landscape and use SD to draw good images, giving more people some thinking.
This model is suitable for the following groups of people: those who have just come into contact with SD and want to immediately start drawing good pictures, those who do not know how to autonomously train Lora, those who want to practice using prompt and drawing beautiful cards, and those who want to use a large model to quickly draw a few good pictures.
This model is not suitable for people who want to train their own Lora model after exploring many models, those who want to use their own training to collaborate with large models for long-term mapping after passing the beginner stage (I still recommend chill here), and those who download a model that is close to overfitting and assume that the effect of SD mainly depends on the large model.
This model will be used in some future domestic courses.
Author's QQ Name: Glacial Acetic Acid (If any group friends know me, hehe)