High-quality models that significantly improve the quality of generated images. Currently, CivitAI is a mature Stable Diffusion model community in the industry, gathering thousands of models and tens of thousands of images with accompanying prompts. This greatly reduces the learning curve for getting started with Stable Diffusion.
Here is a brief introduction to some excellent models on CivitAI, along with application instructions at the end of the article.If you're eager to experience it, click here to go to Anakin and start using the Stable Diffusion XL model immediately.
Model Categories on CivitAI
The models on CivitAI are mainly divided into four categories: Checkpoint, LoRA, Textual Inversion, and Hypernetwork, corresponding to four different training approaches.
- Checkpoint Traiend: A large model obtained through the Dreambooth training method, known for its excellent image generation quality. However, due to training a completely new model, the training speed is generally slower, and the generated model files are larger, typically a few gigabytes. The file format is either safetensors or ckpt.
- LoRA: A lightweight model fine-tuning training method performed on the basis of an existing large model. It involves fine-tuning the model for producing fixed features related to specific people or objects. Notable features include excellent image generation for specific stylistic features, fast training speed, and smaller model files, typically ranging from tens to a few hundred megabytes. It needs to be used in conjunction with a large model.
- Textual Inversion: A method of training models using text prompts, essentially a packaged set of words used to generate fixed features related to specific people or objects. Noteworthy characteristics include excellent image generation for specific stylistic features, very small model files (typically a few tens of kilobytes), but slower training speed. It also requires the use of a large model for effective results.
- Hypernetwork: Similar to LoRA, but with model performance not as good as LoRA. It needs to be used in conjunction with a large model.
Model Recommendations: Checkpoint > LoRA > Textual Inversion > Hypernetwork
In general, combining Checkpoint models with LoRA or Textual Inversion models can achieve better image generation results.
On CivitAI, you can filter and search for models based on different classifications and sorting methods.
After opening the individual model page, you can view the model description and download the model.
Now, let me introduce to you a series of high-quality Stable Diffusion models.
Model Description:Competent in various styles (realistic, original, 2.5D, etc.), capable of generating excellent portraits and landscape images.
Model Description:This model is generally designed for portraits and full-length anime style photos. Fantastic landscapes are quite decent. And it doesn't require kilometer-long queries to get a high-quality result.
3.Realistic Vision V5.1
Model Description:Capable of achieving highly realistic character and environmental rendering, faithfully reproducing real-world styles.
Trigger words: analog style, model shoot style.
Model Description:High-quality 2D, character, and landscape models. Additional loading of VAE models is required to enhance color representation.
Model Description:Can generate models in blind-box style; when using, it is recommended to choose the ReV Animated main model.
Trigger words: full body, chibi.
Model Description:Suitable for drawing urban areas, wilderness, flowers, and a fresh and youthful aesthetic.
Model Description:Icon logo style.
First you need to install the lora extension, and then put the downloaded stampV1_v10.safetensors file into the .\extensions\sd-webui-additional-networks\models\lora folder.When using it, click the lora plug-in tab (the tab may be located under the seed), check the enable option, and select the stampV1_v10 model, and adjust the intensity from 0.6 to 1 according to your needs.For the base model in the upper left corner of Stable Diffusion, select the original default one.
Regarding the strength: If you want the picture to be more like a graphic design, you can make it stronger, such as setting it to 1. But if you want richer elements such as animated images in the picture, you can set it lower, such as 0.7
4.Round animals, balloon shape body
Model Description:Turn animals into a chubby balloon with this lora! it works well itself with a weight of 0.6-1.0; also fine with cg or real checkpoint; pretty good quadrupedal animals, bipedal, not so much.
1.PlanIt! a documentation
Model Description:Instruction manual style.
- scientific diagram,
- scientific illustration,
- concept render,
- architectural plans,
- technical illustration,
- patent diagram
Sepia, Vintage, - for a modern look
Character, person, woman, man - reduce people showing up
Grid, Grain, vignette - depending on the look you're going for
2.Double Exposure for SD 2.x
Model Description:This is the Double Exposure Embedding for SD 2.x.
It was trained with the v2.1 768 ema pruned model on 768px images of people, layered with a variety of surroundings. It's been shown to handle objects and animals good as well.
- Place the file in the embeddings folder of your Automatic1111 installation.
- Trigger the effect in the prompt by writing dblx.
Model Description:This embedding is based on 30 images cooked for a total of 300 steps on base SD1.5: 16 vectors per token, a 0.004 learning rate, a batch size of 6, and 5 gradient steps.
Other Useful Tools
If you want to explore more AI applications and models, you can visit Anakin.ai. There are over 1000 applications for various scenarios, including ChatGPT, Claude, and Stable Diffusion. Go ahead and give it a try!