Overcoming Midjourney Limitations

Hello Visual Veteran

I’m glad to see so many of you have downloaded my latest FREE PDF. I hope you’re having fun exploring new techniques and creating.

If you haven’t seen it yet, you can download it here.

Now, lets get to today’s post. Written by a Guest.

I love seeing how other creatives use AI in their work and I want to highlight one of them today - Marco Morales.

He is a world builder and storyteller creating the world of BubblePunk, a celebration of beautiful and confident plus-size female figures.

This post will show the process of creating an image of a plus-size female figure from start to finish using different AI tools.

I want to give you a sensitivity warning, if you’re not used to such content (nothing NSFW).

Let’s get to it Marco!

Thank you Nejc! To achieve the desired result, I used Midjourney to create the environments and initial poses, then imported them into Leonardo to refine the figure, and finally using Photoshop to compose it all together.

The Midjourney generation:

Midjourney can be quite limited and biased in its treatment of body diversity. It even removes the language needed to describe full figures! Midjourney will usually generate a skinny model or an extremely fat body shape, and nothing in between. Other platforms such as Leonardo or Stable Diffusion provide more variety of body shape and the ability to use more accurate language to describe (and prompt) body-related subjects.

The next step was to upscale the image within Midjourney:

This is as far as Midjourney can take me. Over at Leonardo, I used the IMG2IMG feature to load my Midjourney original, with 0.5 weighting. This is to keep in line with the original image, but leaving some room for improvement:

I then added the same prompt I used in Midjourney, adding better descriptors of the body type I wanted to create. This comes with a warning, as the language I use is deemed too colorful for Midjourney and deserving of a ban! The ‘bad’ words: ‘thicc’, ‘curvy’, ‘butt’ and ‘booty’!

I select Deliberate 1.1 model, which is great for human portraits and gives me a wide range of controls over the figure.

Finally, including also a negative prompt:

The resulting image:

While the overall detail is ok, the faces are not. The next step is to upscale the image.

The upscaled image can be brought into Photoshop for a quick and messy touch-up, and uploaded to Leonardo again as IMG2IMG to integrate the new details in a more natural way. Here is the Photoshop touch-up:

This the Leonardo output with IMG2IMG:

Now I have several long shot scenes and a great close-up scene. I bring all of them into photoshop to compose, pick and choose the details I like most (the flowers) and also to clean up parts that are not right yet: hands and arms.

From here…

… to here

On photoshop I can pick and choose the best flowers and details. Top right corner flower:

Bottom of her dress: (below)

Once I have the final composition, I put it through Leonardo once more, as a high IMG2IMG generation (same prompt, same negative prompt). And upscale (within Leonardo).

I downloaded the upscale and put it through Topaz Gigapixel AI for a touch-up of sharpness and scale-up. My final image (since I am only using it for digital) is 2800 x 2800 px.

I love how immersive Marco’s worlds and workflows are. I always get transported into his imagination and creative mind.

I hope you were able to see some of that in today’s post.

If you want see his whole Bubblepunk collection, you can take a look here.

You can also connect with him on LinkedIn or take a look at his Kepler Files - a Photo Documentary of an Exoplanet.

I hope you enjoyed today’s guest post. I am planning to highlight more creators using AI in their work.

If you have someone you would love to hear from, reply to this message and let me know. I would love to hear from you. 🤗

If you’re looking to have fun with AI, you might like last week’s post.

You can also stop by my website or social media and as always,

Keep creating and don't forget to have fun. ☀