DALL-E 3 vs DALL-E 2: Comparing 18 Prompts Head-to-Head

Soon after OpenAI unveiled GPT-4 final June, they set their eyes on the subsequent item on their roadmap: refining and launching DALL-E 3.

Even just before its release, there have been murmurs that this new model has the energy to rival Midjourney each in creativity and nuance.

Ultimately, when it was officially announced, we acquired our very first glimpse on the AI picture generator of the long term. Individuals (myself incorporated) right away sprung into action and in contrast the new solution to other existing AI image generators in the industry.

But, I’ve by no means genuinely imagined about evaluating its outputs towards its preceding iteration until finally now. There is no query that DALL-E three is a important improvement from DALL-E 2 but how considerably precisely? Let’s uncover out.

Output Comparison

To preserve issues steady, we’ll have DALL-E two on the left and DALL-E three on the proper.

Practical: Individuals

Prompt: a sensible near-up portrait of an elderly particular person that radiates wisdom and character

If we’re going by pure realism, I truly favor DALL-E 2 more than DALL-E three. The elderly guy in its picture has the encounter of somebody who’s been by means of hell and back. Wrinkles, laugh lines, age spots — you just know this is somebody who’s total of wisdom.

It is not that DALL-E three did a negative occupation. I imagined it did fairly properly. Nevertheless, it has a tendency to smoothen faces, which occasionally final results in faces searching like wax figures.

Practical: Landscape

Prompt: a snow-covered alpine landscape with a cozy cabin nestled between the pine trees, smoke increasing from its chimney into the crisp winter air

This is exactly where we see DALL-E 3’s evolution more than its preceding edition. Alternatively of making one thing coherent, DALL-E two seems like it crammed the aspects in the prompt into a single disjointed picture.

On the other hand, DALL-E three produced an picture that is much more correct to the essence of the prompt. It is cozy and fulfills each and every necessity with out overcompensating.

Practical: Historical

Prompt: an picture of new york city that captures the feeling of nostalgia for the 1980s, featuring iconic objects from that era.

DALL-E two fully missed the mark right here. It just produced a bad depiction of a payphone. There are no aspects that scream “New York” or anything at all exclusively about the 1980s.

I’d say that DALL-E three also had a challenging time with this prompt. I would’ve rated it greater but I did not for two causes. A single, why is it in black and white when colour photography was commonplace in the 1980s, and two, why is the subway over ground?

Surrealist

Prompt: a guy-fox hybrid strolling in a surreal dreamscape that combines aspects of a forest, a desert, and a tropical seashore

The landscape itself is fine in DALL-E two, but it did not even try to produce a guy-fox hybrid. DALL-E three understood the context fully and even additional its very own twist in the ultimate output. This is a single of my favored comparisons so far since it properly encapsulates how far this model has come.

I’ve also observed that, although DALL-E three usually smoothens pictures, DALL-E two tends to produce one thing that is rough close to the edges and seems like it really is drawn with crayons.

Architecture

Prompt: a contemporary two-story, eco-pleasant and sustainable residence with solar panels, rainwater harvesting programs, and a green roof backyard

DALL-E 2’s residence plainly suffered from some rendering problems. If you zoom in, you can see that it attempted to produce a balcony on the proper side of the 2nd floor but it turned into a weird window-balcony hybrid. The plants also block considerably of the very first floor, which helps make it challenging to see what the residence truly seems like.

As for DALL-E three, now that is a residence I want to reside in. It is an excellent render of a contemporary eco-pleasant residence. Nevertheless, it does have a dream-like good quality to it, which requires away from its realism.

3D Render

Prompt: a 3D diorama of a fantasy forest with mythical creatures, ancient ruins, and enchanting bioluminescent plants

When once more, DALL-E two missed an essential element of the prompt by creating an artwork alternatively of a 3D diorama. DALL-E three was capable to make an picture of a diorama containing all the context I supplied.

Oil Painting

Prompt: an oil painting of a guy busking in the middle of a bustling city street at evening, with individuals and reflections in the rain-soaked pavement

I like these two pictures since it demonstrates the improvement the model has manufactured in just a yr. DALL-E two happy every thing incorporated in my prompt but it has subpar creativity. It is poorly blended and vague. The reflections are a small bit off also.

DALL-E 3’s painting seems like the DALL-E 2’s output if it was carried out by a skilled artist. The streets search much more comprehensive and alive. It is not ideal (the topic is missing a leg) but it is an really very good try.

Pixel Artwork

Prompt: a pixel artwork scene of a calming Japanese backyard with bonsai trees, a koi pond, and a tranquil bridge

These pictures search like a nevertheless from an previous Gameboy game and its remastered edition. I truly like DALL-E 2’s pixel artwork since of the nostalgia element but, hands down, DALL-E three is far far better. It is vibrant, comprehensive, and steady. The way it seems reminds me of Stardew Valley and Animal Crossing.

Pastiche: Photography

Prompt: a near-up portrait of a musician misplaced in the second, with vibrant colours that evoke the film’s wealthy, with the warm tones of Portra 400

If there is a single factor I genuinely like about DALL-E two, it is how properly it generates near-up pictures. There is genuinely absolutely nothing incorrect with both of these two pictures, but I somewhat favor DALL-E three since of the contrast and lighting.

Pastiche: Artwork

Prompt: a visual homage to Leonardo Da Vinci’s anatomical research such as the intricate specifics of the human entire body with a contemporary twist

What I genuinely dislike about DALL-E two (which led me to use Midjourney in the very first area) is that it tends to produce crayon-like pictures, like the a single you see right here. It is also various from Leonardo Da Vinci’s artwork, which was a essential portion of the prompt.

DALL-E three managed to fix this and produce one thing that seems like Da Vinci’s Vitruvian Guy. It also additional a contemporary twist, as specified in the prompt, by depicting a cyborg alternatively of a human.

Abstract Ideas

Prompt: a visualization of morality

Producing a visual depiction of an abstract idea is usually a hard job and that demonstrates in these pictures. DALL-E two went into a much more concrete depiction exactly where it chose to define morality as an inner struggle in between two sides of your persona. DALL-E three went to a much more abstract street exactly where it depicted morality as a multi-faceted spectrum of various values.

Visualization of Digital Ideas

Prompt: a visualization of the Net as a bodily landscape, with sites and social media platforms as buildings

This a single is not even near. DALL-E 2’s output lacks creativity and the colours mix collectively in a negative way. You couldn’t even inform the buildings apart. On the other hand, DALL-E three did a fantastic occupation at making a society primarily based on the web landscape right now. I also notably like the double-which means of the substantial-pace site visitors going in and out of the web.

Text Generation

Prompt: a information post from the long term reporting on the very first make contact with with an extraterrestrial civilization

A single of the most important issues of AI picture generators right now is text. This is since they perceive texts as shapes with out any which means. Anytime I use DALL-E two, the text usually comes out searching like the Greek alphabet, which you can see right here. 

DALL-E three guarantees to have far better text generation and, for the most portion, it does. It is nevertheless unpolished, but it is undoubtedly an improvement.

Non-Existent Topics

Prompt: an ancient civilization’s undiscovered misplaced city buried deep inside a jungle

DALL-E 2’s option to make an aerial see of the prompt is a head-scratcher for positive. You genuinely can not see any specifics apart from the ruins, which are poorly-rendered to get started with. DALL-E 3’s artwork reminds me of ancient Aztec civilization. It is comprehensive, properly-lit, and has a mythical good quality to it.

Lower Context

Prompt: a flying golden retriever

It is not a very good signal when your AI picture generator can not fulfill a lower context prompt like this a single. Apart from the glaring rendering problems on DALL-E 2’s canine, there is also the truth that it is jumping and not flying. 

DALL-E three was capable to flip a easy prompt into a cute artwork. That explained, I really don’t genuinely recognize the need to have to flip the goldie into an angel, but it adds to the charm of the picture.

Higher Context: Various Factors

Prompt: an alternate background scene exactly where ancient Egypt with sophisticated technological innovation, featuring pyramids with rocket boosters, hieroglyphic-coded computer systems, and cyborg pharaohs, is in a war towards a futuristic Roman legion with electrical spears and mecha horses

Unsurprisingly, DALL-E two couldn’t deal with a substantial context prompt at all. In truth, it looked like it gave up halfway by means of the method. It would be honest to say that its output is a total mess.

DALL-E three, as soon as once more, has blown me away with its precision when it comes to prompts. It did not miss any single line, even when I stored specifying random objects. This one’s a clear winner.

Higher Context: Background Description

Prompt: a neo-noir movie nevertheless of a grizzled detective consuming whiskey in his burgundy-colored chair in the course of a thunderstorm at evening, behind him is a bookshelf filled with books and an ashtray, a single lamp is illuminating a single side of his encounter

At this level, I’m noticing a trend in DALL-E two exactly where rendering problems grow to be much more obvious the longer the prompt is. For instance, the picture over is missing its encounter. Meanwhile, DALL-E three produced the ideal movie nevertheless that accurately depicts my prompt. It is moody, atmospheric, and gritty — just like any other neo-noir movie.

Higher Context: Topic Description

Prompt: an expressive oil painting of a youthful grownup lady of southeast asian descent, with slight curls on her prolonged black hair, who is striving to manage her rage. her fists are clenched as her anger gradually turns into fury. feelings are proven in her encounter and posture. her encounter is gradually getting to be crimson red as she’s conquer with fury.

I’m truly stunned at how properly DALL-E two finished this prompt. It did not miss a single descriptor and I’d say I’m happy with how it turned out. Nevertheless, DALL-E three is just on a total various degree. You can genuinely really feel the feelings coming out of the lady in the image. I also like that, if you zoom in, you can see the brush strokes that is frequent in oil paintings.

My Ideas on DALL-E three

I anticipated DALL-E three to be a far better edition of DALL-E two, but I did not anticipate this kind of a stark distinction in between them to be honest. Seeking at the outputs side by side, it is genuinely evening and day in terms of good quality alone.

It is not just creativity, it also follows by means of with its guarantee of far better nuance and text generation. Missing the interpretation of a line in the prompt would seem to be so unusual in DALL-E three, which is one thing that transpires really frequently even in between Midjourney and Firefly.

That explained, DALL-E three is not ideal. But taking into consideration the strides it manufactured in just a yr, it really is now a viable competitor to other AI picture generators