Google says its Imagen 3 AI image generator beats DALL-E 3. How to try it for yourself


Prompt: Pretty blue shallow ocean with sand.

Sabrina Ortiz/ZDNET via ImageFX

With so many AI chatbots on the market, picking the best one can be challenging. To try and settle the debate, Google DeepMind pitted the leading chatbots against each other and found that users are most impressed by one image generator — Imagen 3. 

Also: I just tried Google’s ImageFX AI image generator, and I’m shocked at how good it is

A report, published on Wednesday, details how Google DeepMind evaluated Imagen 3’s performance against its predecessor, Imagen 2, and leading external models, including DALL-E 3, Midjourney v6, Stable Diffusion 3 Large, and Stable Diffusion XL 1.0, in both human and automatic evaluations. 

The human evaluations tested five quality aspects of the text-to-image generation models: preference, prompt-image alignment, visual appeal, detailed prompt-image alignment, and numerical reasoning. 

In the overall preference category, which measured how satisfied a user was with the image compared to the input prompt, Imagen 3 won with a significant lead over the competition, as seen in the image below: 

GenAI preference performance

Google Deepmind

Imagen 3 performed competitively in the other human evaluation categories, as well as the automatic evaluations, which tested prompt-image alignment (again) and image quality. 

Also: Google’s AI Overviews get three useful updates. Here’s what’s new

“All in all, Imagen 3 clearly leads on prompt–image alignment, especially on detailed prompts and counting abilities; while on visual appeal, Midjourney v6 takes the lead, with Imagen 3 coming in second,” concluded the report. 

“When considering all the quality aspects, Imagen 3 clearly leads in overall preference, indicating it strikes the best balance of high-quality outputs that respect user intent.” 

Sound too good to be true? Here is how you can test Imagen 3 in ImageFX, a tool in Google Labs that lets people create images with simple text prompts.

ImageFX is made available through Google Labs, the company’s platform for testing testing ideas and products. Like all other experiments on Google Labs, accessing the tool is easy and free. 

All you have to do is visit Google Labs and select ImageFX or visit the ImageFX page directly. Then sign in to your personal Google account and start tinkering with the tool. As with any other text-to-image generator, type in a conversational prompt for what you’d like to see rendered. 

Also: The best AI image generators of 2024: Tested and reviewed

A bonus of ImageFX is its fun twist — a prompt interface that includes “expressive chips” that you can use to experiment with “adjacent dimensions of your creation and ideas”, according to Google. Once you type in a prompt, a toggle will appear on selected words of your prompt, which suggests new and amusing ways to tweak it. 

Each generation will render four high-quality images you can enjoy. In my experience, ImageFX even rendered hands well. Hands are often a tricky subject matter for image generators.





Source link