[AI x Side Hustle] AI Image Generation Diary No. 5: Discovering Civitai and Controlling Poses
It's been a while since my last AI development post...! A whole month has passed.
This time, I'm going to develop features related to the poses of the generated images.
Until now, users had to input a skeleton image and the output would follow that image, but now I'm going to make it so the AI can automatically control the poses too!!
And I've actually found a very convenient service for creating image generation AI this time!!
Creating a Dataset
To introduce this feature, all I have to do is create a dataset for the poses! It's simple, haha. The basics of AI are preparing data -> learning from data -> outputting based on learning. I will proceed with the creation based on those basics this time as well.
Meeting Civitai
When it came to creating the dataset, the big problem was where to get the data from... At first, I was looking on Hugging Face and Kaggle, which I introduced before, but I couldn't find what I wanted there.
Specifically, I wanted to collect data on various poses of a single person, such as running or crossing their arms, but there were only poses of multiple people doing something, or sometimes just parts of a body, so I couldn't really be satisfied... That's when I asked ChatGPT if there were any good services, and it introduced me to a site called Civitai. This is quite useful for image generation AI.
🔍 What is Civitai?
A community platform for sharing and exploring generative AI models
Civitai (https://civitai.com) is a platform where you can share, search, and evaluate models and LoRAs for image generation AI, centered on Stable Diffusion. Simply put, it's like a
"material site" and "community" for generative AI. Since it is centered on Stable Diffusion, it's perfect for the AI I'm currently building!
🧠 Main Features
1. Distribution of models such as LoRA / Checkpoint / VAE
On Civitai, anyone can upload and download data for generative AI, such as:
✅ Checkpoint models (the trained AI itself)
✅ LoRA (Low-Rank Adaptation): Lightweight models for additional training on specific styles, people, outfits, poses, etc. (the ones I made last time)
✅ VAE: Auxiliary models that improve the resolution and color of generated images
→ This allows you to teach the AI your preferred art style and expressions.
2. Sharing and evaluating generated examples by users
Each model page contains a large number of images generated using that model (posted examples). When you find a model you're interested in, you can check the generated examples to see if it matches your image.
Also, evaluations (star ratings and comments) and tagging are active, making it easy to find popular models and reliable creators.
3. The search filters are powerful
You can narrow down results not just by keywords, but also by conditions such as the following:
Model type (LoRA, Checkpoint, etc.)
Resolution and source material
Style (realistic, anime, semi-realistic, etc.)
Sorted by rating, popularity, latest, etc.
→ You can accommodate specific needs like "I only want to change the hairstyle," "I want an anime style," or "I only want the skeleton data."
This time, I was able to search for just the skeleton image data.
🛠 You can use it like this!
● Use as training material
As in this case, skeleton data and pose images, which are used as "samples" for LoRA, are very convenient as training material.
● Collect and combine style LoRAs
By combining multiple LoRAs trained on your favorite art styles or specific clothing and hairstyles, you can create a custom model that outputs images to your liking.
● Upload and share your own LoRA
Once you get used to generative AI to some extent, you can upload the LoRA you created yourself to Civitai and share it with other users. I intend to upload the LoRA I made last time!
📌 Points to note
Some LoRAs may contain NSFW (explicit/adult) content, so don't forget your **filtering settings (exclude adult content)**.
With respect to copyright, you must check the license of the original data before commercial use.
When using collected images as training data, cleaning and selection are important.
🧾 Summary: Civitai is a "Generative AI Material Box"
You can explore and share LoRAs and pre-trained models
You can check the finished results in the posted examples
You can also get reference materials for poses and compositions
For those who want to use generative AI in earnest, Civitai is like an indispensable armory. It's easy for beginners to browse and search, so if you're interested, please take a look!
Results of using the dataset
I decided to use the following pose data! There are so many poses, and most of them were exactly what I was looking for, so it's great! However, when I opened the data, there were a lot of gentlemanly ones, so I as much as possible excluded them lol (Take a look if you're interested. They were quite something lol)
To start, I created about 300 skeletal image data files! The images generated as a result look like this
The image on the left still has traces of the gentlemanly poses lol. But the images have become quite stable!
That's it for this time. From here on, I will add art styles and poses, and finally, I will generate characters and introduce a function to generate images of those characters in various poses! The goal is in sight, isn't it! Please look forward to the next one