Midjourney's most anticipated feature for character consistency is finally here. Now, you can use a character reference image to generate characters with consistent details. These are the results I was able to generate from my testing, and I also figured out a workaround to put multiple characters in the same scene with some tricks. In this video, I will walk you through the process of how I generated these results. If you're curious, let's jump right in. So this new cref function works similarly to regular image prompt, except it focuses on the character traits. Midjourney did said that the precision of this technique is limited still, and it won't copy exact dimples, freckles, or t shirt logos. They also said that this feature works best when using the characters made from Midjourney. It's not designed for real people or photos, and they warned that it will likely distort them, as regular image prompts do. But in my test, I actually prefer the real people results, because I think the cref feature is more useful in stabilizing facial features at this point, and not so much in terms of hair and outfit, at least not to my preferred standard for animation or storybooks. these results are great for creating AI influencers or AI models for fashion brands, In this video, I will focus on animation style illustration, as I think most of you watch my videos are creating stories in this style. So first let's try out the cref function in Midjourney. Note that you can use your own images generated from other places as the character reference. But Midjourney said that this feature works best when using characters made from Midjourney images. So here I have an image I created from Midjourney as the reference character. I recommend that once you decide on the character details, you can note down the important features you want your characters to have so that you don't deviate too much from your original character ideas as you continue to generate your images. And here is what I have for my first character, Lily. Lily is a little girl who has big round brown eyes and long hazelnut wavy hair to her waist. Wearing a white puff sleeved shirt and a vibrant yellow flower skirt with suspenders. After getting the character image we like and want to use as the character reference, we can start prompting and see if we can generate the same characters doing different things. I'll walk you through how to generate these images on discord for people who haven't had access to Midjourney Alpha website yet, and I think once you know how to do it on discord, it's super easy to do it on Alpha website, and in the prompt box we type in forward slash. Imagine a little girl picking flowers in the forest in Pixar Animation style. And here I didn't describe too much about the details of that little girl, so we can test some this hero function. You can also use sref to control the style, but here I want to keep it simple with just a text prompt for style. And we can then type in --cref and insert the image URL after the prompt. To get the image URL, you can drag the image directly to the prompt box if the image is easily accessible, or you can click the image open and right click to get the image address, or simply open the image in the browser so you can easily access it from another tab. and just copy paste the link. And after that, we can also use --cw to modify character references. Strengths from 100 to 0. Strength 100 is default and uses all the character details, including the face, hair and clothes as reference and as strength 0 it just focus on the face if you want to change the outfit and hair. Here are the character weight of the same reference image, ranging from 100 to 0 for the prompt. A little girl in Christmas time, Pixar animation style. You can see that the less of the character weight, the less the image resembles the reference character in terms of hair and outfit, but the more it adheres to the text prompt. In this case, the lesser the character weight, the more Christmasy it is. but in this video I want the character to match all the character details as much as possible, so I will just leave this parameter at default without typing anything. And here are the results of our generation. I think it did capture the overall style and facial features of the character, although it's definitely not perfect, especially as you can see the clothing details are not really matching the original image completely, It's still hard to completely replicate all the details, I find it interesting that the generated images did include animals and butterflies, as in the original image, However, we can still see that the consistency controlled by the cref parameter is still a lot better than just using a reference image. So for the details, we can use vary region to edit the image to perfection. first let's upscaled this one. In this panel, click vary region. and select the area you want to edit here. I want to change the suspenders of the dress from blue to yellow like the original one, and let's see if we can get closer to our original character by simply using the cref link and just some simple prompt description. It works. I think U3 is pretty close to the original outfit, so we will upscale this one and add our second character beside it. So here we want to add the second character to the right. At the upscale image panel, Click the pin right icon, and in the prompt box we can adjust the text prompt to two little girls picking flowers in the forest. in Pixar Animation style. After a few tries, I found that in order to generate two girls in the scene with pending, we have to be more descriptive and ask something like this. The girl on the left is wearing a yellow dress. The girl on the right is wearing a purple dress and holding a basket of flowers. If we only use the first sentence and not the second one, most of the time it won't generate another girl in the scene for some reasons. And then we need to switch the character reference to the purple girl. So here we type in --cref and then copy the image link of the purple girl here and send. I think the first image has the basket showing more properly, and the facial expression looks more dramatic. So we will use the first image and fine tune the details using vary region. Same as before. We upscale the image and then click vary region. Here we can try again to select the area we want to edit and see if we can change the proper clothing details. With the flower basket we specify when we narrow down the generation area, and also added the text prompt to be more descriptive about the area we selected in the text prompt I wrote, the girl wears a soft lilac dress with elbow length and puffy sleeves and a purple bow tie at the waist holding a basket of flowers. And then we use the same cref of the Purple Girl image link, and then just click send. This time the girl looks more like the reference image, and with the same logic You can further refine the details to your liking using vary region to change your clothing details, the eye gaze or any other places that you are not satisfied with. If you worry about wasting fast time, you can also edit the details using Photoshop's generative tool, which I mentioned briefly in this video. And here are the images I generated completely in Midjourney. Do you think they are consistent enough? I know they're not perfect, but this is the worst is going to be. Soon enough we will get better and better consistency with these tools. But here is what I got before Midjourney perfects the character consistency function. let me know what you think. And if you're serious about creating the most consistent characters using AI, check out this video for my ultimate Character Consistency Hacks. I will see you soon.
