Celebrating the New Year at home is great, but going back to work is not so pleasant; there's no choice, we have to make a living. So, I found a way to spice up my dull work life, which is to use AI to create cute girls.
Before the New Year, I also tried OpenAI's ChatGPT3, and it was simply amazing. I didn't expect AI to have developed to this extent; it can almost be described as a low-spec version of the intelligent AI in movies, with the downside being that it cannot connect to the internet. But now with new Bing, it can connect to the internet.
Ten or twenty years ago, we talked about "Internet Plus," empowering various industries with the internet, and now, we are likely talking about "AI Plus," using AI to empower or even replace certain industries.
Local Operation#
If your graphics card is older than NVIDIA's GTX10 series, or if your graphics card is AMD, I'm very sorry, but you will need to find other tutorials for installation or use the online operation below for testing.
The installation package comes from a video by Bilibili UP master 秋葉 aaaki titled 【AI Painting】Launcher Official Release! One-click Start/Repair/Update/Model Download Management Fully Supported!. You can also check this UP master's column for common issues and advanced operations regarding the installation package.
Please refer to the video above for installation methods. After installation, please set the launcher configuration.
Select the corresponding graphics card and video memory; insufficient video memory may result in black images or inability to generate images.
-
xattn optimization plan
suggests forcing the use of xFormers to shorten image generation speed. -
Do not use half-precision VAE (no-half-vae)
is recommended to prevent issues with generating images using Lora models. -
Disable NaN check (disable-nan-check)
is recommended to check. -
Skip internal component synchronization (fast startup process)
is optional; if you feel the startup is too slow, you can turn off the startup check, but after disabling this option, please do not install models and extensions from unknown sources.
After clicking One-click Start, wait for the startup process until the following image appears, and the browser interface has opened before proceeding.
This process takes about 5 minutes, please be patient.
After startup, stable-diffusion-webui defaults to Chinese language and the model is set to final-prune.ckpt, which can be used to generate two-dimensional images.
You can select the tags you want from Danbooru Tag Supermarket and then input the positive and negative tags to click generate and see the effect.
Positive tags are: (ahri_1.05), neon lights, v, blue hair, ahoge, hair bow, looking at viewer, 1girl, fox ears, sailor collar, school uniform, blus
You can also use other models, such as Anything, to generate other styles of two-dimensional images; just download the corresponding model in the Model Management section of the launcher.
Generating Three-Dimensional Images#
Dependencies required, please install the dependencies below before using the magic spell:
best quality, ultra high res, (photorealistic:1.4), 1girl, brown blazer, black skirt, glasses, thighhighs, ((school uniform)), (upper body), (Kpop idol), (aegyo sal:1), (platinum blonde hair:1), ((puffy eyes)), looking at viewer, facing front, smiling, laughing, masterpiece, best quality,
Negative prompt: paintings, sketches, (worst quality:2), (low quality:2), (normal quality:2), lowres, normal quality, ((monochrome)), ((grayscale)), skin spots, acnes, skin blemishes, age spot, glan, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry
Steps: 28, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 2861044465, Size: 576x960, Model hash: 2051a5b4f0, Denoising strength: 0.7, Clip skip: 2, ENSD: 31337, AddNet Enabled: True, AddNet Module 1: LoRA, AddNet Model 1: koreanDollLikeness_v10(e2e472c06607), AddNet Weight A 1: 1, AddNet Weight B 1: 1, Hires upscale: 2, Hires steps: 8, Hires upscaler: Latent
If you want to use AI to generate a lifelike three-dimensional wife, you need to use other models and extensions.
You can check this video 【AI Painting】Brand New Fine-tuning Model! LoRA Model Usage Tutorial Plugin Installation NovelAI to install the extension.
As for Lora, my understanding is that it enhances feature training, using a set of images to generate models that reinforce the features of one or several source files. Simply put, if I train with a set of Ragdoll cat images, then use the trained Lora model, input the keyword cat, and increase the model's weight, the AI will likely draw a Ragdoll cat. For training methods, please refer to this video 【AI Painting】Best Character Model Training! Nanny-style LoRA Model Training Tutorial One-click Package Release.
Downloading Models#
To achieve a lifelike effect, we need to download a new model. You can directly download it in the model management page of the launcher; the model is called Chilloutmix. You can also download it from the official page.
If you are downloading from the official site, please place the downloaded model in the launcher folder under
sd-webui launcher\models\Stable-diffusion
This folder stores all the models.
Downloading Lora Models#
The above model can already help you generate lifelike images, but some personalization is still needed.
I recommend the following two Lora models:
This is a model used by many experts, featuring exquisite faces that are indistinguishable from real ones; the example image above uses this model.
This model was created by a Bilibili UP master and can add Ming-style Hanfu to the generated characters.
After downloading the Lora model, you need to place it in the directory under the launcher:
sd-webui launcher\extensions\sd-webui-additional-networks\models\lora
Once everything is configured, you need to modify the settings page for Model's VAE (SD VAE).
You can also use other VAE files to optimize the lighting and details of your work, such as vae-ft-mse-840000-ema-pruned.safetensors.
The downloaded VAE file path is:
sd-webui launcher\models\VAE
Please make sure to save the settings. If you do not set the VAE or save it, the images you generate are likely to be grayish.
After restarting, if everything goes smoothly, you will see the additional options in the web interface.
Check to enable it, then select the corresponding Lora model in the model section. If the model is not found, please click Refresh models below. If it is still not found, please check if the Lora model is in the wrong location.
Then enjoy😊
Online Operation#
If you cannot run it online for some reason, you can also use the online method to try it out.
Please note that most free online platforms have strict resource allocation policies, so they can only be used for trial purposes, and I do not recommend occupying free online platforms for long periods.
The installation package comes from [Share] sd-webui Colab Google Drive Installation Version. Integrated Various Optimizations, Colab notebook address.
Please note that when using the free version of Google Colab, you may be suspended at any time or banned for abuse.
The installation steps are already detailed above. Please note that since the files are stored in Google Drive, your cloud storage space needs to be at least 12GB to accommodate the models.
The default startup interface is in English. You can download the Chinese language pack zh_CN.json here and place it in the folder below.
Then in the web interface settings under User Interface, select the language as zh_CN, save the settings, and restart.
The default installation and default startup model are both two-dimensional style. If you want to generate three-dimensional style images, please refer to the local operation above, then download the corresponding model and place it in the corresponding directory in your cloud drive.
Other Methods#
You can also use the following other online platforms:
Valley of the Gods - Magic Town
Finally, thanks to those who fully embody the spirit of the internet and selflessly contribute.
【AI Painting】Brand New Fine-tuning Model! LoRA Model Usage Tutorial Plugin Installation NovelAI
[Share] sd-webui Colab Google Drive Installation Version. Integrated Various Optimizations, Colab notebook address
vae-ft-mse-840000-ema-pruned.safetensors
Traditional Chinese Hanfu Girl Imitating Ming Style Lora Model
The cited rankings are not in any particular order; once again, thanks to the contributors mentioned above.