Key Learnings: Contrasting Mid-Journey vs. DALL-E 2 for AI Art Creation

Table of Contents

• Default Output Style Varies Between Photorealism and Painterly
• Image Backgrounds Generated More Often in Mid-Journey
• Matte Painting Style Better Captured by Mid-Journey
• Modern Artists Better Represented in Mid-Journey's Training Data
• Specifying Camera Angle Yields Mixed Results
• Content Restrictions Limit DALL-E for Some Applications
• Image Aspect Ratio Fixed in DALL-E but Adjustable in Mid-Journey
• DALL-E Permits Direct Image Editing and Modification
• Controlling Variation Similarity Differs Between AIs

Default Output Style Varies Between Photorealism and Painterly

The first major difference I noticed between DALL-E 2 and Midjourney is that their default output styles tend to vary between photorealism and a more painterly look. Without any prompt modifications, DALL-E 2 tends to generate images that look more photorealistic, while Midjourney's outputs tend to have a more artistic, painted style.

This could be influenced by differences in the training data used for each AI system. The DALL-E dataset may contain more photographs, leading it to bias towards photorealism. Meanwhile, Midjourney may be aiming more towards hobbyist artists and have more painted works in its dataset. However, with the right prompt tweaks, both AIs are capable of producing both photorealistic and painterly images.

DALL-E Tends Towards Photorealism

As seen in the examples, when prompted to generate concepts like stuffed animals or robots, DALL-E 2 often produces images that look like high quality, well-lit photographs. There is a depth and dimension to the outputs that make them appear more true to life.

Mid-Journey Tends Towards Painterly Style

Meanwhile, the same prompts provided to Midjourney result in images with more of an artistic, painted look. There seems to be more interpretation and abstraction rather than rigid realism. The colors are often more vibrant and aesthetics more stylized.

Image Backgrounds Generated More Often in Mid-Journey

Another default difference is that Midjourney is more likely to generate an environment and background along with the main subject, while DALL-E 2 tends to just produce standalone images on blank backgrounds.

For example, a simple "spacecraft" prompt in Midjourney yields spaceships with galactic scenery and landscapes. The same prompt for DALL-E gives a spaceship isolated on a white background.

This means DALL-E may require more specific prompting to situate subjects within environments, whereas Midjourney seems primed towards fuller scene generation.

Matte Painting Style Better Captured by Mid-Journey

When trying to emulate the aesthetic of matte paintings used in films, Midjourney appeared to capture the desired style much better than DALL-E.

The Midjourney output had a clear painted look with visible brush strokes, color variation, and a better sense of scale and perspective for a matte painting.

The DALL-E attempt was darker and lacked the defining matte characteristics, seeming more like a photo composite.

Modern Artists Better Represented in Mid-Journey's Training Data

Tests trying to mimic the styles of specific modern artists had mixed success between the AIs. Midjourney managed respectable interpretations of artists like John Berkey or John Harris. But DALL-E struggled to emulate their aesthetics.

However, both systems produced plausible outputs when prompted with older artists like Van Gogh and Cezanne. This suggests Midjourney may have more contemporary artists still represented in its datasets, while DALL-E's training data likely leaned towards more classical works or just lacked coverage of modern creators.

Specifying Camera Angle Yields Mixed Results

Attempts to specify exact camera angles and perspectives had middling results in DALL-E. Simple directives like "close up" or "overhead view" failed to reliably yield the desired framing.

Angles like "low angle" resulted in some, but not all, images captured from below. And a "3/4 view" prompt generated front-facing shots despite requesting a specific orientation.

So while DALL-E appeared to interpret angle instructions better than Midjourney in some cases, it still struggled with consistency between outputs for the same phrasing.

Content Restrictions Limit DALL-E for Some Applications

Due to AI safety considerations, DALL-E 2 has strict content restrictions that prohibit generating certain imagery like violence and weapons. This unfortunately limits its usefulness for applications like video game concept art, where depicting guns, soldiers, and action scenes may be required.

Midjourney does not seem to enforce similar restrictions, so offers more flexibility in generating a wider range of imagery, regardless of potentially sensitive content.

Image Aspect Ratio Fixed in DALL-E but Adjustable in Mid-Journey

The images produced by DALL-E are locked to a set square aspect ratio with no option to modify or adjust this. Midjourney, on the other hand, allows specifying portrait, landscape, or square orientations.

Post-processing would still enable cropping and resizing DALL-E outputs. But the ability to directly generate different orientations in Midjourney streamlines workflows that require images tailored to specific formats or layouts.

DALL-E Permits Direct Image Editing and Modification

A standout feature available in DALL-E but not Midjourney is the ability to directly edit an existing image by painting over sections of it with new text prompts.

So a generated person could have their outfit or hairstyle revised by deleting those elements from the image and substituting descriptive text of the desired changes. This enables iterative refinement of images that isn't possible in Midjourney currently.

Controlling Variation Similarity Differs Between AIs

Both DALL-E and Midjourney can provide variations of an initial image prompt. However, controlling the degree of similarity seems to work differently between them.

Midjourney offers adjustability via "variation" parameters but wide-ranging changes are easier to produce. DALL-E's variations stay closer to the source by default but lack controls for greater divergence.

So Midjourney may suit early concept exploration with looser sampling while DALL-E allows focused refinement of a more resolved idea.

FAQ

Q: Which AI art generator is better for photorealistic output?
A: DALL-E tends to default to a more photorealistic style, so it may be preferable for photorealistic AI art.

Q: Can Mid-Journey emulate a matte painting look?
A: Yes, Mid-Journey seems to capture a matte painting aesthetic better than DALL-E based on test results.

Q: How do the AIs compare for generating modern artist styles?
A: Mid-Journey appears better able to reproduce styles of modern artists based on the training data used.

Q: Can DALL-E edit images directly?
A: Yes, DALL-E has an image editing feature that allows directly modifying parts of an existing image.

Q: Which AI makes controlling variation similarity easier?
A: DALL-E generally produces variations more similar to the original, while Mid-Journey varies more loosely for the same prompt.