Layers

When people generate images, they often want to adjust just one part of the result. Maybe the composition is right but one object is the wrong colour, or it's sitting in the wrong spot. Having to regenerate the whole image to fix one small thing is a blunt tool, and you can easily lose the parts you liked.

The idea behind Layers was simple to say and hard to do well: break the image into discrete layers, so each part can be worked on by itself.

The goal was to turn one flat image into layers that people can move, scale, rotate, reorder and edit with AI. The team started with a particular model in mind to power the split.

Fraser, one of our engineers, worked with Keenan, our product manager, to quickly build a proof of concept in the browser. That gave us something real to react to early. From there, I joined to help refine the implementation and change the design, building on the structure we already had in the editor.

Some interaction details are really hard to explain in words. How much should a row lift when you pick it up? How quickly should the other rows get out of the way? For this I drew on Josh Puckett's series about “dialling in” details (Interface Craft). A small HTML tool with a few sliders makes it easy to dial in the details you'd struggle to describe.

So I built single-file HTML prototypes for different parts of the design. Each one is a little tuning bench: a stage on one side, the real UI in the middle, and a panel of knobs to play with. Both of these are the actual files, so have a go.

Once a tool was configured the way I wanted, I copied its settings and gave that configuration to Cursor Cloud Agents. The agent did the implementation in our codebase and opened a PR for review.

I really liked how this split the work. Making the design artifact and implementing it stayed separate, so the prototype could stay a quick, playful thing and the PR could be reviewed on its own merits. It also freed me up to spend my time with collaborators on the bigger questions: choosing the right problem, and finding a good solution that builds on the structure we already had.

It came together quickly. We took it around the team for feedback, and then kept refining. There was a lot of going back and forth, catching small bugs and taking in suggestions, and the feature got better with each round.

Layers launched on Envato.com in August 2026. You can split an image into separate layers, move and rearrange them, and edit one layer with AI while leaving the rest alone. It detects up to 16 layers, and you can export them as PNG layers or as a layered PSD for Photoshop.

For me, the bigger result was the way of working. Tuning details in small HTML tools, then handing those settings to an agent to implement, meant the design decisions arrived in the codebase intact, and I could spend more of my time with the team.

Once we'd shipped the feature, I made this AI video to share on my social media to promote it. It's a great way to engage users and get people excited about trying the feature, both internally and externally.