We tried EVERY AI Image Generator to figure out which is best
Science & Technology
Introduction
In a recent exploration of AI image generation models, we set out to compare various popular tools in a bid to determine which one produces the best results. With numerous requests to assess the effectiveness of models in generating photorealistic images, alongside those of a more artistic nature, we developed a tool that allows us to hit multiple image generation models simultaneously. Currently, we have support for models from OpenAI, Stable Diffusion, Black Forest Labs, and MidJourney.
The Experiment
To commence our testing, we decided to use prompts that would challenge the models on both realism and creativity. Our first prompt was to generate a photo of “a lifeguard on a beach.” This provided a platform to analyze the models' capabilities in rendering humans and complex environments. Notably, the Black Forest model produced the most photorealistic image, featuring a feminine lifeguard, whereas the others leaned heavily towards male depictions. MidJourney and Stable Diffusion's outputs resembled action figures rather than realistic humans.
Next, we challenged the models with a prompt depicting “a rocket mid-launch.” Here, MidJourney stood out significantly with a realistic representation, placing it ahead of the others, confirming its proficiency in rendering high-particle activity scenes.
Moving onto a more artistic prompt, we tested “an impressionist painting of the French countryside.” Unsurprisingly, there was no clear favorite, yet MidJourney's output exhibited details closely aligned with what one might expect in an art museum setting. On the contrary, the outputs from Stable Diffusion and DALL-E were less convincing, lacking the traditional nuances synonymous with impressionism.
Shifting gears, we explored creativity further by requesting “a minimalist illustration reminiscent of mid-century poster style, depicting a man at a diner.” MidJourney excelled here, creating a composition that echoed the time period, while the results from DALL-E and Flux missed the mark.
We explored more fantastical themes with prompts like “a pack of woolly mammoths trekking through snow,” where MidJourney again led with a more lifelike representation. Further inquiries, including “a closeup of a person drinking out of a soda can,” saw Flux struggling with clear hands, while MidJourney offered a more coherent albeit still imperfect image.
Finally, we sought to test the whimsical capabilities of the models by asking for “an alien on the planet Zeta Reticuli commuting to work.” Here, we noted that all models interpreted the prompt diversely. MidJourney provided a polished cinema shot, while DALL-E’s output felt reminiscent of early AI-generated images.
Conclusion
Through the various tests, we found that Flux performed exceptionally well in the realm of photorealistic images, while MidJourney emerged as a strong contender overall—capable of balancing creativity and realism effectively. Stable Diffusion showed potential but had discrepancies in some outputs, and DALL-E ultimately fell behind the competition. Moving forward, we hope to delve deeper into these models and provide more specialized recommendations based on their strengths.
Keyword
- AI Image Generation
- Photorealism
- Black Forest Labs
- MidJourney
- DALL-E
- Stable Diffusion
- Impressionism
- Creativity
- Artistic Representation
FAQ
Q: Which AI image generator produced the most photorealistic images?
A: The Black Forest Labs model was found to generate the most photorealistic images in our tests.
Q: How did MidJourney perform in the tests?
A: MidJourney consistently produced high-quality images and excelled in both photorealistic and creative prompts, often taking the lead overall.
Q: What was the weakest model in the comparisons?
A: DALL-E was considered the weakest model, with numerous outputs that did not meet the expected quality compared to the others.
Q: Did the models handle creative prompts well?
A: Yes, models like MidJourney and Stable Diffusion performed better in creative outputs, while Flux struggled to meet artistic expectations.
Q: Is there a tool available for public use to test these models?
A: Currently, the tool we developed for testing is not in a shippable state, but if there is sufficient demand, we may consider making it available for public use in the future.