Hands-on with Google’s experimental AI

Generative AI has both impressed and dismayed people with its ability to turn text prompts into images or even more text. But earlier this year, Google began testing AI designed to generate things even when the inspiration for what to type into that prompt box is missing. As part of the Google Test Kitchen, ImageFX and MusicFX aid users in suggesting what to ask for in order to generate off-the-wall ideas for images and even music .

Related

11 annoying tasks Google Gemini will soon handle for you

Gemini 1.5 Pro will soon be able to answer questions about the world around you using video, among other key updates from Google I/O.

The tech itself isn’t terribly estranged from Google’s more widely known Gemini. In fact, ImageFX uses the same text-to-image diffusion model as Gemini and Google Lens. But what the experimental programs are designed to do is power more ideas and creative thinking by listing alternative keywords to use in the prompt.

We asked Google’s MusicFX to create some songs for us, then asked ImageFX to generate an album cover and even a band poster. But did the Google Test Kitchen leave a bad taste, or are the tools the future of AI?

Related

5 cool things Google’s Gemini AI can do on your Pixel 9

The new Google Pixel 9 phones have some exclusive AI features.

What is MusicFX and ImageFX?

An experiment or something more?

MusicFX and ImageFX are experimental AI currently being tested by Google, along with similar options like TextFX. Google’s FX tools are generative AI for when you don’t know how to write the prompt. The web-based software is designed specifically for experimentation and exploration, rather than the productivity focus of Gemini inside Gmail, for example. Both are free to try in the U.S. at the Google Test Kitchen.

MusicFX turns text prompts into short songs, up to 70 seconds long. The experimental AI also helps users write the prompt, suggesting what to say even before typing anything into the box. Once you add a prompt, the software changes keywords into drop-down menus called Chips. For example, in the prompt, “Country music inspired by crows, played on the guitar,” country, crows, and guitar all offered multiple suggestions to try. With just a few clicks, I could turn that original prompt to, “Blues music inspired by owls, played on harmonica.”

Related

10 Gemini Live features I can’t wait to try

Google’s AI sounds more human-like, but what, exactly, is Gemini Live capable of?

Similarly, ImageFX is an AI image generator that helps refine text-based prompts. It’s a large language model powered by Imagen 2, the same technology that Gemini uses. While Gemini can already generate images, ImageFX suggests changes to the prompt, turning keywords and phrases into drop-down menus. These so-called chips are designed to give the user more ideas and better refine the result.

In fact, ImageFX doesn’t even require an initial prompt — simply hitting “I’m feeling lucky” randomly generates a prompt for you. “Dreamy, pastel landscape, soft lines, gentle colors, fluffy clouds, rainbow mountains, minimal,” can become, “Mystical neon portrait, angular forms, bold colors, dramatic clouds, jagged mountains, ornate,” all without tapping the keyboard.

Related

Gemini Live is here, allowing voice conversations with Google’s AI

It’s available now if you’ve got Gemini Advanced.

MusicFX generated short jingles

Soulless, but not terrible

When I first started testing MusicFX, I quickly figured out what the music generator can and cannot do, a mix of abilities that was both at times relieving and disappointing. First, I wasn’t able to get MusicFX to generate vocals, though I did get some do-re-me’s when I asked for a cappella. And, much to the relief of artists everywhere, you can’t ask the program to replicate a certain artist. Sorry, but MusicFX won’t be cranking out any Taylor Swift songs anytime soon.

MusicFX’s creations are limited to 70 seconds long, at the moment, but you can toggle on the loop option for it to seamlessly replay itself. The default is for a 30-second song, but you can adjust the length by opening up the settings menu.

Related

SearchGPT explained: What it is and how you can be the first to try it

OpenAI has long been rumored to be working on a competitor to Google Search, and now it’s finally here.

Prepared to listen to music as disastrous as the three-fingered, melted face portraits of early image generators, I was surprised when I clicked play to find the tune wasn’t horrible. It was music that I could imagine playing in an elevator, or the background while waiting on hold. After the first result wasn’t terrible, I generated a few more, trying multiple genres, speeds and instruments.

After a while, the songs the software generated started to all feel similar to each other (though in hindsight, perhaps I shouldn’t have asked for so many country songs). While the clips are short, there’s no sense of structure, like a chorus or verse, but there seems to be shorter beats that repeat themselves with slight variations. While I wasn’t cringing, I also wasn’t humming or tapping along to the beat either. The music lacks the soul and emotion of the songs that normally I can’t help singing along to.

Related

ChatGPT Free users can now generate DALL-E 3 images, although only two per day

You can now generate images from text in ChatGPT without the need for a subscription.

At times, the software wasn’t able to listen to exact instructions. For example, when I asked for music played only by acoustic guitar, it still created a tune with multiple instruments. I can’t see MusicFX making any Billboard hits, but I can see it generating the background music for video ads and commercials. But, with the copyright controversy around generative AI, it’s unclear if the resulting image could, or even should, be used commercially.

The best part of MusicFX, however, is those drop-down chips designed to unleash more ideas. In my opinion, one of the best features of AI is the off-the-wall randomness, which occasionally turns out to be a spectacular idea. Using the different suggestions, it’s fun to try something new or take one idea and lead it even further. The way it comes up with off-the-wall ideas like “bubbly, optimistic cyber pizza party music at the underwater arcade” is quite fun to experiment with, although I was disappointed when I Googled that suggested prompt and found it had spit out the same suggestion many times before.

The best way to customize the results, however, is with DJ Mode. With this option, each part of the prompt has a slider, so you can turn up the upbeat tempo or turn down the classic rock feel. This way, you have more control over the final results. As ideas occur to you, you can add them to the list, or use the suggestions at the bottom. DJ mode has yet to gain the download and share features, however.

With MusicFX we generated a country song, a psychedelic jingle, and, in DJ mode, a classic rock with acoustic instruments.

Related

Does Apple Intelligence actually stand a chance in the AI race?

If Apple can flesh out its in-app tools, it has a shot at standing out against the competition.

Some ImageFX results were horrifying, but others were impressive

The software suggested prompt tweaks to take the photo in new directions

Google / Pocket-lint

Naturally, after coming up with a few different AI-generated tracks, I had to make an album cover to go with it. For that, I employed ImageFX, a tool powered by Imagen 2, the same sub-set of Gemini that generates graphics. Like MusicFX, it uses chips to suggest adjustments to the prompt, from the style to what’s generated.

The first prompt I asked for resulted in the three-armed, white-eyed clown-like musician that will probably now haunt my nightmares. Reminded of how difficult it is for AI to replicate a human form, I adjusted my prompt and was surprised at how quickly I found something that I liked. The AI nailed the style that I was going for, which was reminiscent of a vintage circus poster.

What was most impressive, however, was that the AI was able to handle text. The AIs that I’ve worked with previously have never been able to correctly add words, creating gibberish and misspellings, even when I just asked for a simple “happy birthday.” If I told ImageFX what words to include, however, it added those words correctly spelled. It’s not perfect — when I didn’t specify what words to add to the album cover, it added letter-like shapes to a part of the design obviously intended for text. But, it’s more impressive than the text on an image I’ve attempted to generate with ChatGPT.

Related

Tim Cook reveals when ChatGPT will be added to iOS 18

In Apple’s latest earnings call, the CEO confirmed that ChatGPT integration will arrive soon.

Here are a few of the images it generated:

Is MusicFX and ImageFX the future of Gemini?

If there’s one feature I want to see in Gemini, it’s the Chips

Generative technology, especially that that attempts to replicate art, demands questions about what place, exactly, the technology has in our future and how it affects actual human creatives. If MusicFX is any indication, I can see AI-generated songs as hold music, elevator music, or the forgettable background music to a social media video. I can’t see myself jamming in my car to anything that the tool has created so far. But, as it stands, MusicFX is experimental and could make huge leaps ahead as it progresses.

Another question that needs to be addressed with every machine learning platform is where the training data comes from. Google has not disclosed where it found the music to train the system. However, a report from Billboard suggests the company used copyrighted music inside its training set. With lawsuits ongoing over the use of copyrighted images in training data, legislation could play a significant role in whether MusicFX makes it out of the Google Test Kitchen.

Related

Gemini and Google Workspace can help you be more productive… most of the time

Google’s Gemini is a pro at summarizing Google Docs and emails, but things get a little quirky when it comes to Sheets and other Workspace tools.