Midjourney has surprised the world with its new Image Prompting feature in its Version 4. This has completely changed the way people and especially artists are gonna use this tool. It has opened up opportunities never seen before. In this article we are gonna explore the potential and possibilities of this new weapon in our arsenal.
How is Image Prompting in V4 is different than other Midjourney versions?
In earlier versions of Midjourney, the image input was considered more of a style rather than an image. The A.I. would try to emulate the style of the uploaded image and generate an image mainly based on your words. But this time, it’s completely different. It takes the shape and subjects of your input and creates arts using the image as a basis and a skeleton. That gives someone more control than ever on how your image is gonna look even before you have typed any prompt. Previously the only way anyone had to add their unique touch and creative juices to the outputs of Midjourney was to alter the images in software like Photoshop, Procreate etc after they have been generated. We can retouch them, change the colors, merge and mash them up with other images and so on. But everything could be done after the image has been generated. And the only way we could guide the image generation process was by writing prompts and trying to fine tune them. Sometimes the outputs were predictable, sometimes not and it would take lots and lots of variations to get to the image we actually visualized. But not anymore. Now with v4 we can input an image with all kinds of images including basic shapes, complex diagrams, photographs, portraits & many more and guide the image generation process any way we so desire. Now graphic designers, painters, sketchers all can add the A.I. magic to their works and make their work better and unique in every way.
How is Image Prompting in Midjourney V4 is different than Image to Image in other AI image art generators?
MJ v4 image prompting is not like image to image other AI based image generators. It isn’t like stable diffusion where you train the AI with a model locally and create awesome portraits of someone and also put them in different scenarios. It is also not like Dall-e where you remove part of the image and prompt the a.i. to generate something there from your imagination. Instead, Midjourney image prompting reinforces its already existing approach to art. Here you guide the A.I. to create images while using your images as a base whereas Midjourney keeps its approach towards art forms intact. It looks similar at first glance but here unlike using the uploaded image as ‘seed’ in image to image process, the AI ‘understands’ the components of the image prompt and builds on that. It’s more like an image to text to generation process. If you want to do something like train your own model or inpaint images with the help of a.i., then the open sourced Stable Diffusion or Dall-e might be the best bet for you. If you love the art style of MJ and and want to guide the AI according to the will of yourself them MJ would probably be the best choice for you.
How to Input an Image as A Reference to The Prompt in Midjourney?
First things first. Let’s learn how to put an image link into midjourney for the program to recognise it. No, it’s not drag and drop method, but it’s also not far from it. For those accessing discord via computer, it’s very simple. You upload an image in any discord server or drag and drop an image into the server, hit enter, and your picture is uploaded now. Now you click on the image, click ‘original image’ button on the bottom right corner. The image opens up in a new tab. All you have to do is to copy the url from the tab and paste it into your discord server with the bot after /imagine, write your prompt and hit enter. Here is a beautiful illustration of the steps by Kris Kashtanova. You should pay her a visit at kris.art.
For mobile phone it gets a bit tricky. Although most people figured out the computer part, I got asked a lot by my friends about How to input and image as a referenece in mobile via discord? Many thought it was not possible. But surely there is a way. You can’t go as easy as copying the url directly but you can do it in another way. After uploading the image into your discord server you click the image. That opens up the image. Now you click the share button from the top right corner and choose one of the options from the list. I generally use the note taking application. There the link to the image opens up. Cut or copy the link from there, paste into the message after /imagine, type the prompt and generate.
Creating Portraits using image to image in Midjourney
People have kind of love-hate relationship with this right now. Some people absolutely love this as they now can get their portraitsand other images made via Midjourney in different interesting styles, while others hate it as no matter how hard they try, they can’t seem to get it look like themselves. So, I decided to dive deep into this. I used various different images to experiment with this feature. Here are some of them.
<image> male –v 4
<image> male, smoking cigarette –v 4
<image> male, doing yoga –v 4
<image> male portrait in studio, wearing a grey suit, red tie, spectacles –v 4
<image> full body shot, a person walking down the street, walking with his dog, people passing by –v 4
Here I used simplest prompts to influence the image as little as possible with the words and let it work with mostly the image. Still, it couldn’t produce any image which would seem like the portrait of the original person. The third image from first set of generation seems closest to the original person but still not the same. I tried out different variations and rerolls of the prompts but couldn’t get an image like the original.
<image> bright white angel, white wings, real photo –v 4
<image> real male, working in a computer, cyberpunk style –v 4
With this set of images I realised that the AI confuses the gender if I don’t specify it. As you can see in this example, it has created a female version of the character. With different styles it’s still not the same person at all.
 Then I decided to include more than one image of a person in an image. therefore, I inserted 3 different images of the same person from different angles to get the best possible result and tried out different prompts. I inserted these two images along with the previous image. After many tries only one set of generation gave somewhat similar looking dress and face that too where I gave no word input. But it was definitely not the same person. All those prompts, variations and rerolls gave only one result where you could at least recognise the face that too if you look very closely. 
<image a> <image b> <image c>Â –v 4
<image a> <image b> <image c> middle aged man, wearing red tie, working in computer, cuberpunk –v 4
<image a> <image b> <image c> man wearing armour, kawai, pixar eyes, renderman, artegerm, 8k –v 4
What I observed is that when you give it some other style that’s little bit far from being real humans, the results somewhat becomes acceptable. My theory is, when we try to create realistic images, we tend to prefer results closest to the original. But as this AI doesn’t produce such results, we don’t really like them much. But when we are trying to get the images in a different style like cartoon or animation or any other style where we don’t expect real images but similar ones, we accept them more readily even though they are not the same. Midjourney seems to be always altering the images slightly to avoid deepfakes. That can be a good or bad move depending on your preferences. With that being said, I don’t think we can use MJ v4 to create different portraits of ourselves but we can definitely use them for a fun experience or even use it to get consistent results for things like graphic novel and all.
Here is a great tutorial by Kris Kashtanova about getting a consistent character with Midjourney V4
P.S. Many of you might be wondering why I only used male models for my experiments and examples. That is because Midjourney seems to ban any and every image I tried to include as an image prompt. All of these three images were banned due to somehow being Potentially NSFW Image Prompt. I really don’t know how this happened but since all of my image prompts including female models were termed as potentially NSFW content, I refrained from using any images of a woman at all.
Creating Interior Designs Using Midjourney A.I.
Probably the biggest impact of this img 2 img feature is gonna be on the design industry. This is where Midjourney v4 is much much different from any other of its competitors. In this first example I am going to use a sketch of a kitchen and try to turn it into a fully colorful 3d model for use. I used the nijijourney bot, which is now in beta test, to create a sketch of the interiors of a kitchen. Interior designers that are skilled in such drawings can now easily create designs just from their drawings. Instead of spending days in creating a 3d model from their imagination, they can now create it in a matter of minutes and try out different ideas that they previously couldn’t due to time constraints.
In these examples I shall try to keep the prompts as simple and as minimal as possible as the purpose of this tutorial is give you the structural idea about how all these works. I’ll make another detailed tutorial where I talk about how to specify each part of the prompt to generate exactly what you want and fine tune the results. Now let’s jump in.
<image> real kitchen, hd photograph, shot from nikon camera, unreal engine, with top and bottom shelf, red textured drawers, interior design 3d modelling, blender –v 4
<image> real kitchen, hd photograph, shot from nikon camera, unreal engine –v 4
<image> real kitchen, hd photograph, shot from nikon camera, unreal engine, teal colored chimney, red textured drawers, interior design 3d modelling, blender –v 4
As you can see, even with fairly simple prompts the midjourney v4 gives great results already. All you have to do is to guide the algorithm. This is a great tool for interior designers to showcase their creations in colors and in a realistic way pretty easily. Although it would be already enough for most of the usages, I wanted something even more realistic. And we all know what gives us most realistic results in Midjourney, it’s remaster. As Midjourney v4 doesn’t yet support remaster, it was a little bit tricky but not much hard at all. I just created a remix of the v4 image and changed the —v 4 prompt to —testp —creative and the result it produced is just amazing.
 
<image> real kitchen, hd photograph, shot from nikon camera, unreal engine –v 4
<image> 3d design of a real kitchen, hd photograph, shot from nikon camera, unreal engine –testp –creativeÂ
<image> 3d design of a real kitchen, hd photograph, shot from nikon camera, unreal engine –testp –creative
<image> 3d design of a real kitchen, hd photograph, shot from nikon camera, unreal engine –testp –creative
Next I created a sketch of a study room via the same method. Then created the v4 version of it which already gives pretty great results. Adding little bit of ‘Remaster’ salt again creates a great looking study room.
<image> open space, long corridoor, bookshelf, sofa, –v 4
<image> real photograph and 3d render interior concept design of open space, long corridor, bookshelf, sofa, –testp –creative
Creating Architectural Designs Using Midjourney A.I.
Now going for creative possibilities in architectures. Here I used nijijourney to create a sketch of a petrol pump. Then modified the prompts step by step to finally create different 3D versions of a petrol pumps based off of the initial image and as you can see all of them are built over the first image that I put in.
<image> real 3d render petrol pump, at night, light shine –v 4
<image> isometric diorama 3d render petrol pump, at night, light shine –v 4 –no background
<image> isometric diorama 3d render petrol pump, at night, light shine –testp –creative –ar 9:16 –no background
<image> isometric diorama 3d render petrol pump, at night, light shine –testp –creative –ar 16:9 –no background
Here I created a shopping mall using an image generated in nijijourney.
<image> shopping mall, futuristic design, change color to red –v 4
<image> eal photograph, real building, 3d render of architecture shopping mall, new york city, futuristic design, change color to red –testp –creative –ar 16:9, Remix
Does it Really Follow Your Image Prompt or Would it be the Same Without the Image?
In this example I am gonna try out the same prompt without the any image and with different images. Whenever I change the number of islands or design in the drawing, the same thing happens in the generated picture too. I used the simplest pen drawing possible to highlight this feature.Â
 
floating island city –v 4
<image> floating island city –v 4Â
Then I changed the prompt to suit my taste. Although they looked crappy at first, they got better and better as I went with variations and remixes.
<image> floating island city, extremely realistic builiding, architecture, VRAY render, 4k, realistic details –testp –creative –ar 16:9
Among all the generations, these two looks best for me. I hope this clears your confusions about if this really follows your input image or not.
My Take on This
The v4 update is a huge milestone not only for Midjourney and A.I. generated art enthusiasts but also for many other people working in other creative fields. I have started seeing some professionals, who weren’t interested much in AI generated arts and its possibilities before, are now suddenly showing interests and trying to explore ways to include AI in their own workflows. Although you can not create your portraits by this yet and maybe never will be as they would always slightly alter the image to avoid deep fake, they might develop it to the point where you can choose which altered version to be kept and used in all kinds of images and all kinds of things. Thus you might be successfully be able to generate different images in different scenarios and scenes to create something like graphic novel or photo stories with same consistent character and with your own images or some model of your choosing. But the most interesting part I think is the ability to turn your sketches into 3d models. This opens up a huge world of possibilities for designers and architects. You can have 3d models of your sketches now within minutes that too in different variations and from different angles. Soon the aspect ratio options will also be available in v4 giving people even more creative freedom to make designs for various personal and professional works.
Image to Image and Its effects on Copyright
Initially the only way to change or modify an AI generated image in Midjourney via human input was altering or modifying the image after it has been created. Now you can do it even before any image has been generated. By using your images, drawing or photos as prompts you can create results that no one else can generate. Thus the created arts become truly your own. That gives people even more solid grounds on the issue of copyright. Although there is still a cloud of confusion whether ai generated images can be copyrighted or not, with human created copyrightable images being used as prompts, gives even more grounds for these images to be copyrightable. With that being said this A.I. tech is still relatively new and already moving at super high speed. I learn something new every single day. I shall keep you all updated with new things happening everyday. Keep creating…
