0:00 Did MidJourney just solve the character consistency problem? They released their long awaited only reference feature and with that they replace their character reference parameter. Only reference can be best thought as a system of saying put this in my image. It works for characters, objects, vehicles or non-human creatures. It should work with personalization, stylization, style references and mood boards too, which we will test in a second. Is it a step forward for consistent characters? Let's explore together. To be able to use omnic consistency, ensure that you selected version 7 under your settings. For this test, I will rely on style row for now. And I will turn off my personalization for now. To use omni consistency, all you need to do is uploading an image first and dragging
0:46 and dropping this image to this new omni reference section on the right side. I think best way to test this feature is using my own photo. I uploaded my photo and I'm going to drag and drop this to omni reference section. Okay. Now let me put myself somewhere. Let's say cinematic photo of a Victorian detective. You will realize that now next to my photo there is an omni reference strength slider. It's also called omniate and default will be 100. I think we can start with the simple prompt and default value of 100 for omni weight and let's see what this will bring. So this is my first ever test with omni reference. Okay, this is pretty impressive. So, especially here, I see a quite lot of likeness to my original photo. You can see comparison on the screen. This one as well, it
1:33 looks like me pretty much. This is definitely step forward in comparison to d- CF character reference parameter. So, that was pretty much hit or miss, but this worked out quite well actually. And then you can feed these results to clean and make something cinematic out of it. Let's try another camera angle and see how it handles the different camera angles. So I'm going to use the same prompt with same omni reference and this time I'm going to ask cinematic high angle photo. I'll send this and I'm going to also send side profile photo. Let's try bird's eye view photo. All right, this is pretty impressive. Resemblance is quite high and it managed to gave me high angle shot as well. So,
2:20 bravo, good job for midjourney. Very happy with this result. Side profile photo looks very impressive. And bird's eyee view shot where I'm about to fall from a building. Okay, this didn't necessarily quite worked out fine. Um, but we know midjourney's problem with distant faces. So, it's not a big surprise. Now, let's try another character. I directly dragged and dropped to only reference section this particular photo of a woman sitting in a cafe. So, it's not like a perfect close-up shot. And she has an outfit that I will try to replicate her face and outfit alto together. And I'm still staying with the style raw for now. So, bear with me. I still stay with photo realism. So, I'm going to say it's a fashion photo shoot, but she's
3:05 meditating. Very interesting. Okay, let's hit this with first of all default 100 value. So, in the press release, they said if you want to keep the clothing consistent, you need to aim for something around 400. So let's try this. Okay. So here in these results I see a little bit of confusion. So for your reference this was the original image and I see these results doesn't necessarily look like her. So distant face with 100 default value struggled little bit but I boosted number to 400. One thing I immediately realized is clothing got much more consistent to original image. So skirt, white shirt, brown belt. If you want to replicate
3:52 exact clothing of your reference, you need to increase omni weight to around 400 value. The face, on the other hand, it's not bad, but I was expecting something a little bit more of a closer resemblance. Of course, it's not like a close-up shot where mid Journey can easily render the Omni reference, but still it's a good test. It is not possible to drag and drop multiple images as a reference to this section. So therefore, I'm guessing multiple consistent characters is not here yet. David said it's possible to hack it, but it's not fully tested. So I won't go into that direction, but I definitely want to test the object reference. For this job, I'm drag and dropping this complex object. Looks like an
4:38 astronomical clock, but like a magical one. almost like a time traveler device. It's a complex object. I wrote cinematic photo of a steampunk time traveler holding a multiverse meta time clock. And I use this object as a reference point. And looking at the object, it definitely resembles especially here it resembles quite a lot. But here I see the sphere on top of the clock is little bit too huge. Here it looked good with 400 omni weight. I got much better results especially here when I dropped the omni weight to around default value of 100. This resembles quite a lot what I originally fit as an omni reference. This is also very good. So I would recommend you to be careful with omni
5:24 weight. In my quick test I realized that object consistency worked much better with default value. Let's do another object consistency test for people who love to use AI for their product photography jobs. So, I have this can of a Coca-Cola and I would like to have a product shoot with this. And for this one, I won't make it super easy for mid journey fashion photo shoot with a female model in a studio. She's holding a product. So, I didn't specify Coca-Cola because then maybe it would pull it out from its training data. I just wrote she's holding a product. That's it. All right. So, we got mixed results here. We have some struggle both with hand and the product itself. However, this is completely off here.
6:10 The can writings are off. And here we got little bit closer, but scale is wrong here. So, I think this is still not perfectly ready for product photography, especially for pixar perfect e-commerce type of things. I will also quickly increase omni strength to again 400 and see what happens this time. Meanwhile, while this is generating, I have another vehicle reference. It's a very specific 10 wheels Mustang GT, the famous LNR car with very unusual look. Cinematic photo of a vehicle in golden hour. Okay. Next to the beach. Increased omni weight, but results didn't necessarily improve. For
6:55 object consistency, 10 wheels Mustang didn't work with default omni weight. It still gave me regular four wheels. The color and brand is correct. So, thumbs up for that. I'm going to again increase the Omni strength even more this time to 600 level. Run again. Same issue with the higher on new8. It got the brand right and color right, but it struggled with the nuances of the product. But then if you give little bit of clue to midjourney that you want a Mustang with 10 wheels additional to object reference then results get definitely improved but still not the 10 wheels. I just feel like midjourney is hitting a limit here of an imagination and midjourney team says this feature works with non-human creatures as well and of course what's
7:43 the best way to test this with your own dog. Say hi to Looney. Among four images of my grid, one resembles little bit looney. The other three it's really not even close. I will try this with higher values as well. After increasing omniate my grit is definitely improved. So I see one good result here. You will realize it also copied the yellow bandana here. So good job for midjourney and face looks quite okay here. The resemblance is really high. So very good result except five flex. Typical AI is struggling with details but face consistency looks very very good. So I'm happy. So another takeaway point if you're going to use it with your pets or
8:30 maybe non-human creatures you definitely need to increase the omniate. Okay. Now let's go back all the way to my original image and this time I'm switching to midjourney aesthetics and let's take this around 300. I would like to use new experimental parameter d- exp. From what I understand this is closer to stylized. Therefore I would expect a little bit more artsy results in combined with stylized value. with default midjourney aesthetic stylization, it managed to keep the face reference really well and got maybe a little bit more dreamy, maybe got little bit more, let's say, colorful. And this is with exp. So I'm using this parameter for the first time.
9:16 And I can definitely see that here we got little bit more, let's say, somewhere between photo realism and illustration. This looks like it's from a graphical novel. Let's use same original reference and with 100 default value but personalization on. As expected, my personalized style gravitates towards realism and the reference works perfectly fine. Ancient Japanese aesthetics and I'm going to use my face reference with this particular moodboard style. And let's switch to midjourney aesthetics a little bit more artsy and bump up this to around 250 I would say would be sufficient. Wow, this is so cool. It kept the face reference
10:01 and kind of inserted that into the art style of the mood board. You can do so many cool stuff with this. But I think among these four, I like this the most because it match the art style and face consistency together. So here's the last thing. And let's test star reference together with omni reference. You will realize that I have this quite artsy star reference. It's quite illustrative, strong reds and whites. And let's use this together with my face reference. Okay, I like this more because after bumping up the style weight and specifying that I want a digital art, I got much more arty looking results, much more illustrative. So, key takeaways, the face consistency work really well, especially for photo realism. If you add
10:46 a close-up photo of your subject's face, with distant faces, things get little bit more complex. If you want to copy and paste clothing of your subject, I would recommend you to bump up omniate value around 400. With object consistency, I would recommend you to keep omniate around 100 default value because I got worse results when I increased it. it still struggles with unusual looking vehicles. If you're going to ask an unusual object consistency, it makes sense to mention that in your prompt to strengthen your message for the model. Uh non-human subject references worked much better when you increase the omni weight to around 400 or higher values. And it definitely works with stylization and mood boards and style references, but
11:31 sometimes it defaults more into photo realism, especially with face. So in order to prevent that, you need to increase weight of your style. So you can use style weight for this job. Hopefully this video was truly helpful for you. Don't forget to give a thumbs up and subscribe for more in-depth tutorials. If you want to learn more about future of storytelling, click
