We'll guide you through a step by step creative process for multiple character consistency.
Step 1
First you'll need to train your character LoRA's. In this example we will work 3 characters. For more information on training consistent character LoRA's, see the following articles here in the knowledge center:
Training a Character LoRA
As you can see in the following examples, the datasets show diverse and varied poses and emotions.
Step 2
After curating your dataset, focus on captioning your images during the training process. In this case you'll want to use a "unique token," and we recommend an approach you might see often in the generative AI community. Choose a name for your character, and remove all the vowels leaving only the consonants. Briefly describe some details in the image. For a deeper dive into captioning check out our Knowledge Center article here:
Guide to Manually Captioning Your Dataset
As you can see, the output images are very consistent with the input. Playing with captions and prompts can maintain even more consistency, and it depends entirely on your goals. You can also train clothing items in the same method as characters, and compose them together to build a model. See our articles here for more information:
LoRA Compositions
For this workflow; to create a through line with your characters you can blend them together to create a connection through the style of the blended model. We chose Picturesque Anime Landscape. It just adds a slight nuance to the characters trained style.
Step 3
Next you will start to add prompts to guide your character. A light touch is often all you need, and here you can see we use Prompt Expressions to further guide the result. For more information on using prompt expressions please read this article:
Using Prompt Expressions
Also, and most important, the images we are going to use in the final creation need to be Upscaled and then subsequently Isolated to remove their backgrounds. It is important to do the Upscaling first. Don't worry at this stage about any odd details, because we will finish assembling everything in the canvas.
Here are the 3 outputs we chose from the three blended models we made.
We also used the Liminal Spaces model to create a spooky tree-filled background, which we will then Send to Canvas.
Step 4
You can download your images and then upload them your new canvas project, or you can select them directly from your Scenario library. We'll start with the first layer, the spooky silver haired girl.
.jpeg?width=688&height=320&name=GDbrD3_XsAAWKPG%20(1).jpeg)
Now upload your second character. Because this is meant to be dramatic, like a comic book issue cover, We are going to give the knight a bit of a drop shadow and blur it to make it less stark.
We will do the same for the red haired character. As you can see, her dress was somewhat degraded when we removed the background. However, that is not a problem, because we can fix it with in-painting. I horizontally flip her layer and give her a shadow first.
Next, in-paint her dress by first switching the Model, so it is the model that was originally used to create the image and so will follow and match her style.Then mask out the area you want to change. If we wanted to only make a small change, we would adjust the Influence (under Reference Image). However, in this case, the default 25 is good.
Feel free to do a lot more of this kind of editing, changing small aspects and dialing your creation in. Explore and create until you are satisfied!
