Noobs guide to character consistency in Image models
The invention of image generative models has transformed the way we create and interact with digital content. It has led to intriguing use cases, in gaming, film making, and photography. However, creating images of the same character is still challenging.
In this blog, I will tell you how to create a character, and then create images of the same character using LoRA training with no technical knowledge.
Images generators like Midjourney, Flux or Stable Diffusion are great at generating creative and realistic image. However, its hard to make them spit image of the same person.
Ability of create images of the same person is called character consistency.
Many people use face swapping technique to achieve character consistency, however this method fail severely to generate creative and complex images. Also, its tough to match the skin color of the face to the rest of the body.
There are other effective ways to achieve character consistency which involved training or fine tuning your own image model. Some of these techniques are Dreambooth and LoRA. I prefer LoRA over all other method because it is easier, faster, and cheaper to train. It also requires less images to fine tune.
However, you need multiple images of the same person to train a LoRA. I have covered this in another blog.
Well, we can solve this issue if we have a few headshots of same AI human. Thankfully, it is easy to generate multiple headshots of same person from a single photo. And we can generate that single photo using AI.
After we have a couple of headshots, we can fine-tune a LoRA to create more complex images for your AI influencer account on Instagram or TikTok.
And we have just the right tool that can do it in a few clinks inside your browser, and it works on your phone too.
Full disclaimer — I am the creator of this tool, and this tool is not free.
Creating an AI human
It is a three step process and takes about 10 minutes to setup, and then 30 minutes to train during which you can close your laptop and enjoy your coffee or visit your favourite club.
Steps are pretty simple-
- Specify and choose a model of your liking.
- Generate and choose 5–10 good headshots through AI.
- Train the model
Now we will go through each steps in the web-platform. Read the image captions for more details.
Specify and choose a model- Choose a name that is unique, so that AI can associate that name with your model without confusion. Choosing a common name (like celeb names) will simply confuse the AI, and produce poor results. The names can be any random words like XUOSOS, or whatever.
Generate and choose 5–10 good headshots I have automated the flow to generate unlimited headshots of your selected influencer. Not all photos will be good. That is why we train a LoRA, to get more reliable and creative photos. Choose the images that resembles your chosen model in the step 1. Make sure you have a variety of clothing, face expression, and background. Select 6–12 photos.
Train and generate Now click on Finish to start fine-tuning your own model. It takes around 25–30 minutes to fine-tune. You can close the browser. Generation process is same as any text to image model. Make sure you use the model name (in this case - Lyra2) in the prompt.
Challenges It takes a bit of creativity and practice to write a good prompt to generate pretty images. I will add some customisation template to ease the process. Till then, be creative and put those hard work.
Future Creating an influencer is just the first step. I have the following plans. Suggest me if you have more. Find me on Twitter @saquiboye.
- Video reels - convert the images into captivating videos. Right now, you can accomplish this using any Image to video tool.
- Make your influencer endorse certain product- like a shoe, suitcase, cloths, etc.
- Make short films using your consistent characters.
Create now - TheFluxTrain Gallery for Lyra2-
Thank you.