Today, we’re going to discuss an incredibly interesting topic in the field of AI painting: why can Midjourney easily handle various styles with a 5.2 model, while Stable Diffusion requires us to switch between countless models? π§
In Stable Diffusion, creating an image might require you to choose from models like Realistic, Anime, 2.5D, etc. Want a specific style? Then you need to add a LORA model. For even more optimized results, we might even have to use ControlNet and VAE models, which feels like an endless puzzle game! π΅π«
Although it seems complicated, each model and LORA is designed to provide us with a more personalized and customized creative experience. Behind these models, a significant amount of time and resources have been invested; for instance, the SD1.5 version cost as much as $600,000 to train, with training time measured in tens of thousands of hours!
However, the official purpose is not to use these models directly to generate images, but to pre-train them and improve the foundational level. This way, even with just a regular consumer-grade graphics card, you can create a model with a personal style by slightly training on the official model.
This is why you see so many customized models on the market. But here comes the problem: due to the lowered barriers for training models, many models vary greatly in quality, and their universality and compatibility are concerning. Additionally, some platforms have launched incentive programs, prompting everyone to release models in bulk, which not only leads to model homogenization but also exacerbates the tight disk space situation.
To address these various issues
Today, we recommend some very useful versatile models to help you reduce the difficulty of choice.
Juggernaut XL Series
Basic Information
Source: civitai
Author: KandooAI
Type: checkpoint
Update: February 24, 2024
Version: V9+RDPhoto2-Ligthning_4s
Base Model: SDXL_Lightning
Recommended Parameters
Resolution: 832*1216+
Sampler: DPM++SDE
Iteration Steps: 4-6
CFG: 1-2
Proprietary Keywords by the Author
Photography, Wildlife Photography, Car Photography, Food Photography, Interior Photography, Landscape Photography, Hyper-detailed Photography, Cinematic, Movie Still, Mid Shot Photo, Full Body Photo, Skin Details
Features
This model is considered photography-level, covering a wide range of styles, including architecture, animals, cars, food, interiors, landscapes, and more. In generating characters, unlike other models with generic AI faces, the skin details and expression portrayal are exceptional, with the only downside being that it may not perform well with Asian faces.
Download Link:
https://pan.baidu.com/s/12PpbJcLNT4Qhbag6L1P2jQ?pwd=p43d
DreamShaper XL Series
Basic Information
Source: civitai
Author: Lykon
Type: checkpoint
Update: February 20, 2024
Version: v2.1 Turbo DPM++SDE
Base Model: SDXL Turbo
Recommended Parameters
Resolution: 832*1216+
Sampler: DPM++SDE Karras
Iteration Steps: 4-8
CFG: 2
Features
This model emphasizes versatility, with the author specifically using the symbol β (infinity) as its unique identifier. The author dislikes the closed nature of MJ and vows to compete with Midjourney and DALL-E, hoping to spread the spirit of open source and freedom. This model generates images very quickly, with high quality, whether photo-realistic or artistic. Due to its versatility, it excels in realism, 2.5D, and anime styles as well.
Download Link:
https://pan.baidu.com/s/1fSZKFuEWWSUrcqu5X2owoA?pwd=cegz
LEOSAM’s HelloWorld XL
Basic Information
Source: civitai
Author: LEOSAM (a well-known model trainer in China)
Type: checkpoint
Update: February 11, 2024
Version: HelloWorld 5.0 GPT4V
Base Model: SDXL Lightning
Recommended Parameters
Resolution: 1024+
Sampler: DPM++ 2M Karras, Restart
Iteration Steps: 20-30
CFG: 6-8
High Definition Repair Algorithm: ESRGAN 4x / 8x_NMKD-Faces_160000_G
High Definition Repair Multiplication: 1.5
High Definition Repair Steps: 8
High Definition Repair Intensity: 0.3
Proprietary Keywords by the Author
film grain texture, analog photography aesthetic
Features
A well-known model trainer in China (ε η²), this version is marked by GPT-4V and has undergone significant fine-tuning in the fields of science fiction, animals, architecture, and illustration. This version offers more diverse and dynamic character poses and compositions, creating visually impactful images. It has utilized a large film dataset as training material and has specifically enhanced the film texture, which can be triggered by proprietary keywords. It also enhances the expressiveness of themes such as science fiction, horror, and animals, making designs for mecha and similar themes more compelling.
Download Link:
https://pan.baidu.com/s/15RtcvCpIrlFyKHRZxKCoAA?pwd=tkg6