CASE STUDY: Midjourney Portrait Photography
This guide will show how to leverage Midjourney v7 to create more realistic photographs of models. We will start by creating a moodboard, describing this to a general character description, add it to an image generator for a portrait of a model with good skin texture and another image generator providing better prompt adherence. The pipeline will have Midjourney as main focus for tools, but we will also integrate ChatGPT, Cosmos, Reve, Midjourney and Adobe Photoshop.
- Cosmos.so → Moodboarding
- ChatGPT.com → Image recognition & description
- Reve.art → Providing better skin texture compared to Midjourney as initial photograph
- Midjourney.com → Iterating tool for photographs based on the initial photograph
- Adobe Photoshop → Compositing
1. Cosmos
Create an account on Cosmos.so
After succesfully registering, you log in through https://www.cosmos.so/login
On the main dashboard, press the ‘Create’ button. Here you will click on ‘New Cluster’. Name this as the name of your moodboard, then press on the white ‘Create’ button again in the pop-up tab.
After creating a new cluster, go back to the discovery page. In the middle search bar, enter the theme and subject you want elements around. In this guide we go with the theme of ‘Y2K’ and I search for an aesthetic I can make my moodboard around. This is the output that you receive.
Once you hover over an image, you can press the ‘+’ button to add it. You have to make sure that it will be added to the correct board. Double check this in the top left of your image to see the name of your preferred moodboard. In this case it’s in the ‘Y2K’ moodboard.
If you don’t seem to find the results you were looking for. You can always use the assisting tools that are provided on the right of the search bar.
Here you can choose a dominant color. Once chosen, your images will revolve more around images that have this color in their composition. This tiny tool can be helpful when you lack inspiration.
Once you have made a board that’s big enough according to your needs. Click on the right profile image icon (the circle with a gradient color) to see the overview of your moodboards that have been created. Once you click on the moodboard we just made. You get a general overview of all elements that have been picked.
To get a view of the complete moodboard, make sure to press CTRL + Z + MousewheelBack to zoom out. Once you can see all the items. Make a printscreen of the moodboard. For windows this is done by using the ‘Snippet tool’ or for Mac users press SHIFT + COMMAND + 3.
Save this picture to your preferred destination on your computer.
2. ChatGPT
Search for https://chatgpt.com/ and create an account.
After creating your account, make a new chat. Now copy the image we saved or drag and drop it in the chat box. Ask ChatGPT “Describe me this character in 1 paragraph.”
This is all the output we needed from ChatGPT. Let’s continue to the next tool, Reve.
3. REve
Visit reve.art and click on ‘Go to app’. Once logged in, you can hover over the navigation bar to click on ‘Account’. Here you can see your current credits left. Reve provides users with 100 credits to use freely. If you have less than 20 credits, they will update it back to 20 every day.
Now we hover back on the nav bar to go to the ‘Create’ tab.
At the bottom of the page, we see a grey box. Here we can enter or paste our prompt. This is the prompt we got as output from ChatGPT. Re-read it if you are not sure of the prompt and tweak it to your means. It does not have to be specific with settings, more like a good description of your ideal model to be photographed.
“This character embodies a bold, surrealist take on Y2K cyber fashion, blending nostalgic early-2000s tech aesthetics with futuristic rebellion. With their spiked red hair, heart motifs, and retro sunglasses, they channel punk and rave subcultures while donning a red suit that contrasts sharply with their unconventional headgear a tech visor seemingly constructed from handheld devices or faux screens. The look is theatrical and hyper-stylized, almost cartoonish, echoing the rise of AI-generated fashion and avatar identity. This figure feels like a digital-native rebel from a dystopian metaverse stylish, defiant, and unmistakably post-internet.”
With these settings, we will get an image with 3:2 ratio, 4 images in total with image seed on so there will be variety. This is the output we receive.
This will be the image we will continue with. Save it to your computer and let’s continue to Midjourney.
4. Midjourney
Visit midjourney.com and click on log in.
It is recommended to use your google account for more ease of use.
Once logged in, make sure that you click on the settings button.
Have the settings adjusted to the image below.
In order to use Midjourney v7, you will need to go through a survey named ‘Personalization’ by Midjourney where you will be shown 200 images with a choice of A and B, you pick the image you like and you progress through the survey. Once completed you will gain access to version 7. It is set to version 6.1 by standard, to make this guide useful we will need the latest one, v7.
Image size: To change the aspect ratio of the image generated.
Model
- Use Raw Mode for image effects on your images.
- Choose your default Midjourney version.
- For quick prototyping, use Draft Mode.
Aesthetics
- Experiment with Stylization to influence your image with how artistically generated it should be.
- Use Weirdness for unique quirks.
- Variety is used for more diverse outcomes.
More options
- You can select the default GPU speed.
- Use Stealth Mode to keep your image generation session not visible to others on the Midjourney feed.
After this is done, we will click on the icon to the left of the search bar.
After the image has been uploaded, drag it to the ‘Omni-reference’ tab.
Now we can start prompting for a few try-outs.
Using the same prompt from ChatGPT that we used in Reve, we can see that Midjourney gives images that are exaggerated and unrealistic. The skin texture is missing and the headwear gives an unrealistic look. Reve will therefore be the choice to use providing the initial image.
“This character embodies a bold, surrealist take on Y2K cyber fashion, blending nostalgic early-2000s tech aesthetics with futuristic rebellion. With their spiked red hair, heart motifs, and retro sunglasses, they channel punk and rave subcultures while donning a red suit that contrasts sharply with their unconventional headgear a tech visor seemingly constructed from handheld devices or faux screens. The look is theatrical and hyper-stylized, almost cartoonish, echoing the rise of AI-generated fashion and avatar identity. This figure feels like a digital-native rebel from a dystopian metaverse stylish, defiant, and unmistakably post-internet.”
Midjourney v6 without input reference, based on prompt only.
Midjourney v7 without ‘Omni-Channel’ reference, based on prompt only.
For a more detailed way of Midjourney prompting, visit Guide to Midjourney
In general sense this is the prompt structure used by many in Midjourney v7. The 2/4/2 method.
2: Subject, / Setting
4: Detail / Detail / Detail / Detail
2: Atmosphere / Emotion
For this guide it is in the best practice to set up your prompts as having your portrait shot in close-up, the mm of the lens, the perspective and the studio environment.
Midjourney v7 with ‘Omni-Channel’ reference.
Now to get to the full potential of Midjourney v7 we will make more use of the OmniChannel settings. We re-use our prompt and adjust it with extra information, this info is based on our style reference.
Start prompt
Dutch-angle 28mm flash portrait, shot from ground level up, black studio photography
Updated prompt
Dutch-angle 28mm flash portrait, a grid, shot from ground level up, black studio photography of a man showing different emotions, side view, front view, smiling.
with setttings of Aspect ratio 3:4, Omni-weight set to 400 we get this as output.
OMNI CHANNEL
This is a new image reference system in Midjourney named Omni-Reference or Omni-Channel. It has the possibility to duplicate items from an image and put it in a new image. This can work for characters, objects, vehicles or non-humanoid creations.
Simplified, you can think of the system as “put THIS in my image”.
To control the strength of the omni channel, use the slider we have shown from the settings bar
The important part here is the weight/amount of the strength. Since there is no indication what is good or bad, the default has been set on 100 as strength.
For example:
If you want the style of your image from photo to anime, you should put the weight/slider to 25.
If you want the face of the character to be more visible or that the clothes are preserved, you should use 400 as weight. This is the sweet spot in context of this guide.
Best practices
- Your text prompt is important. You need to your preferred scene as it is, write your prompt accordingly and make sure you write down your scene with additional details that are not showing in your reference image.
- Reinforce the preferred style: If you want your generated image in a different style than the refererence given, mention your style at the beginning of your prompt and at the end. Like “Painting of a man with red hair painted by Monet”. Additionally, use the style reference tab (where you put an image for style reference) and lowering the omni-weight for example, therefore you would need to be more specific in your text prompt to reinforce the physical appearance you want to preserve.
- While using the Omni-Channel, you can only upload one image. But let’s say you have an image with multiple characters in it, you can try to mention this in your prompt to have them (the characters) copied to the new image.
- A recurring problem may be that freckles, logos and watermarks may not be copied correctly to the new image, there will be an offset most of the time.
- Experiment with other parameters: If you use high stylization weight or —exp value in your prompt you may want to put the Omni-reference weight to a high weight as well. All these parameters try to dominate in influence.
EXPERIMENTING
Old prompt:
Dutch-angle 28mm flash portrait, a grid, shot from ground level up, black studio photography of a man showing different emotions, side view, front view, smiling.
New prompt:
Dutch-angle 50mm flash portrait, a grid of 2×2, shot from ground level up, black studio photography of a man showing different emotions, side view, front view, smiling.
Simplified, you can think of the system as “put THIS in my image”.
To control the strength of the omni channel, use the slider we have shown from the settings bar
I noticed it would not use the Dutch-angle, a certain angle of shooting photos. So I tweaked the prompt more like this and started over:
After a second try, I selected the third image and asked to upscale it with the ‘Upscale > Creative’ button.
I was still not conviced about this output so looked at other examples that were generated
I did the same with the selected picture and pressed the ‘Upscale > Creative’ button again.
💡Still not convinced I adjusted the prompt with additional lighting settings. Used our first initial model as the character reference and the image above as the image reference to make me a new image.
Old prompt:
Dutch-angle 50mm flash portrait, a grid of 2×2, shot from ground level up, black studio photography of a man showing different emotions, side view, front view, smiling.
New prompt:
Dutch-angle 50mm flash portrait, a grid of 2×2, shot from ground level up, black studio photography of a man showing different emotions, lots of club lighting, side view, front view, smiling.
To come to an conclusion, we can see that the Omni-Channel released in Midjourney v7 is very powerful to bring consistency in images. We will use this image to composite a front page mockup of a fashion magazine in Adobe Photoshop.
5. Adobe Photoshop
Once you open Adobe Photoshop, click on File > New
Paste the image on the newly created canvas. Select the edges that are missing with the ‘Select’ tool and use the Generative Fill function to fill up the space.
“Expand the background colors and the suit of the model.”
Now this is our current composition. To separate the model from the background, we use the ‘Object Selection Tool’ and select our subject/model.
If the model is selected as shown in the image above, we press CTRL + J to duplicate the selection to a new layer.
Now double check to see if the model has been rendered out of the background by disabling the ‘eye’ icons of the first two layers.
Check them again so the eye icon is showing on all layers, and select all three layers together. Press CTRL + T to go into the Transform tool and resize the image on the canvas to get a good composition of the model.
As this is the setting for our final composition, we will add a new layer below the newly duplicated layer of our model. The layer we create now will be used as a text layer to show behind our model.
Select the second layer and press the “+” button.
After following these steps, you are able to mockup designs for magazine covers, advertisements and other visual directions in a quick and efficient way.