xAI's new image model finally nails text 🎨

Good morning. xAI dropped a new image model that fixes the thing every AI image tool has been bad at, actual readable text. Luma rolled out a feature that lets you go back into a finished AI image and change one piece without wrecking the rest. And a guy who runs an eyeglasses shop in China made a heaven video so good the whole internet assumed a studio built it. Which of these are you clicking into first? Hit reply and tell me. We cover all three below.

xAI launched Imagine Image 2.0 on August 7, live now inside Grok as the new Quality Mode on the web and the iOS and Android apps. And it actually renders text correctly. Small labels, dense layouts, multi-line captions, the stuff that's come out as alien squiggles on almost every AI image tool for years finally comes out sharp and readable. It also plans layout more like a designer would, so busy multi-part visuals hold together instead of turning to mush.

The bigger deal is editing. Instead of regenerating a whole image and praying, there's a magic wand that changes only the spot you point at and leaves everything else untouched. You can select a precise area to swap, pull a clean subject out with a transparent background, feed it up to five reference images at once, and resize a single approved image into different aspect ratios. There are also prewired templates for repeat jobs like product shots, headshots, icons, and merch, so you fill in your inputs and skip building the setup each time.

Where it actually ranks: xAI says Image 2.0 is second in the world for both making and editing images, though that's xAI reading the public leaderboards, and the model sitting in first is OpenAI's latest image model. So call it a strong number two, not the new king. For creators, readable text means you can make thumbnails, titles, and posts with words baked in and not fix them in Canva after. Precise editing means you stop regenerating twelve times to change one thing. One catch, the developer API is still rolling out, so for now this lives in the Grok app.

Luma rolled out a feature called Luma Layers that treats a finished image as a stack of editable pieces instead of one baked-in picture. Feed it a poster or a product shot and it splits the thing into its parts, background, subject, headline, logo, shadow, each on its own layer with a transparent background. From there you move, hide, replace, or restyle any single piece without touching the others. It's the Photoshop layers idea, except the AI does the separating for you in one step.

Why this beats normal AI image editing: the usual problem is that changing one thing means regenerating the whole image, and the whole image comes back slightly different, new lighting, shifted layout, a face that's now a stranger. Layers kills that. Turn a silver watch gold, swap the headline text, move the product two inches, and everything you didn't touch stays exactly as it was. For anyone running a campaign where fifty images need to share one consistent look, that's the difference between an afternoon and a week.

This is where AI images are heading in general. The first wave was about generating a cool one-off. This wave is about control, keeping the parts that work and changing only what needs to change, which is how actual production has always worked.

A sub-one-minute AI video showing a traditional Chinese vision of heaven blew up across TikTok and X this month, over 5 million views in under four days. Towering celestial palaces with curved roofs, a white jade bridge stretching into a sea of clouds, red-crowned cranes gliding between temples that float in the mist. No dialogue, no subtitles, just the imagery, styled straight out of classic myths like Journey to the West. Most people who saw it assumed a film studio or a big 3D animation team made it.

It was one guy. An eyeglasses-shop owner named Qiguang, born in 1991, who taught himself AI image and video generation over two years in the downtime between customers. The video wasn't one prompt and done, it took hundreds of discarded attempts. He conceptualized each shot, generated high-quality still frames one at a time, tweaking prompts over and over to get the architecture and lighting right, then animated those stills into motion. Reporting pegs the actual cost at a few dozen yuan, call it single-digit dollars.

Stuck on a tool or have an AI video question? Reply, I’ll do my best to help.

Created something cool? Send it over and I might feature you. 📺

📫 Want to get your tool featured or advertise to our readers? Let's chat → [email protected]

Thanks for tuning in!

-Kevin