Monthly Archives: July 2024

AI stuff I need to see in Photoshop

AI/ML

…and other creative imaging tools, stat!

Google Research has devised “Alchemist,” a new way to swap object textures:

And people keep doing wonderful things with realtime image synthesis:

“How To Draw An Owl,” AI edition

AI/ML, Illustration

Always pushing the limits of expressive tech, Martin Nebelong has paired Photoshop painting with AI rendering, followed by Runway’s new image-to-video model. “Days of Miracles & Wonder,” as always:

Meta releases SAM 2 for fast segmentation

AI/ML

Man, I’m old enough to remember rotoscoping video by hand—a process that quickly made me want to jump right out a window. Years later, when we were working on realtime video segmentation at Google, I was so proud to show the tech to a bunch of high school design students—only to have them shrug and treat it as completely normal.

Ah, but so it goes: “One of history’s few iron laws is that luxuries tend to become necessities and to spawn new obligations. Once people get used to a certain luxury, they take it for granted.” — Yuval Noah Harari

In any case, Meta has just released what looks like a great update to their excellent—and open-source—Segment Anything Model. Check it out:

You can play with the demo and learn more on the site:

  • Following up on the success of the Meta Segment Anything Model (SAM) for images, we’re releasing SAM 2, a unified model for real-time promptable object segmentation in images and videos that achieves state-of-the-art performance.
  • In keeping with our approach to open science, we’re sharing the code and model weights with a permissive Apache 2.0 license.
  • We’re also sharing the SA-V dataset, which includes approximately 51,000 real-world videos and more than 600,000 masklets (spatio-temporal masks).
  • SAM 2 can segment any object in any video or image—even for objects and visual domains it has not seen previously, enabling a diverse range of use cases without custom adaptation.

Neural rendering: Neo + Firefly

3D, Adobe Firefly, AI/ML

Back when we launched Firefly (alllll the way back in March 2023), we hinted at the potential of combining 3D geometry with diffusion-based rendering, and I tweeted out a very early sneak peek:

A year+ later, I’m no longer working to integrate the Babylon 3D engine into Adobe tools—and instead I’m working directly with the Babylon team at Microsoft (!). Meanwhile I like seeing how my old teammates are continuing to explore integrations between 3D (in this case, project Neo). Here’s one quick flow:

Here’s a quick exploration from the always-interesting Martin Nebelong:

And here’s a fun little Neo->Firefly->AI video interpolation test from Kris Kashtanova:

AI in Ai: Illustrator adds Vector GenFill

AI/ML, Generative Fill, Illustration

As I’ve probably mentioned already, when I first surveyed Adobe customers a couple of years ago (right after DALL•E & Midjourney first shipped), it was clear that they wanted selective synthesis—adding things to compositions, and especially removing them—much more strongly than whole-image synthesis.

Thus it’s no surprise that Generative Fill in Photoshop has so clearly delivered Firefly’s strongest product-market fit, and I’m excited to see Illustrator following the same path—but for vectors:

Generative Shape Fill will help you improve your workflow including:

  • Create detailed, scalable vectors: After you draw or select your shape, silhouette, or outline in your artboard, use a text prompt to ideate on vector options to fill it.
  • Style Reference for brand consistency: Create a wide variety of options that match the color, style, and shape of your artwork to ensure a consistent look and feel.
  • Add effects to your creations: Enhance your vector options further by adding styles like 3D, geometric, pixel art or more.

They’re also adding the ability to create vector patterns simply via prompting:

Photoshop’s new Selection Brush helps control GenFill

Adobe Firefly, Generative Fill

Soon after Generative Fill shipped last year, people discovered that using a semi-opaque selection could help blend results into an environment (e.g. putting fish under water). The new Selection Brush in Photoshop takes functionality that’s been around for 30+ years (via Quick Select mode) and brings it more to the surface, which in turn makes it easier to control GenFill behavior:

Magnific magic comes to Photoshop

AI/ML, ControlNet, Relighting

I’m delighted to see that Magnific is now available as a free Photoshop panel!

For now the functionality is limited to upscaling, but I have to think that they’ll soon turn on the super cool relighting & restyling tech that enables fun like transforming my dog using just different prompts (click to see larger):

Realtime face editing with LivePortrait

AI/ML

I wish Adobe hadn’t given up (at least for the last couple of years and foreseeable future) on the Smart Portrait tech we were developing. It’s been stuck at 1.0 since 2020 and could be so much better. Maybe someday!

In the meantime, check out LivePortrait:

And now you can try it out for yourself:

tyFlow: Stable Diffusion-based rendering in 3ds Max

3D, AI/ML, ControlNet

Being able to declare what you want, instead of having to painstakingly set up parameters for materials, lighting, etc. may prove to be an incredibly unlock for visual expressivity, particularly around the generally intimidating realm of 3D. Check out what tyFlow is bringing to the table:

You can see a bit more about how it works in this vid…

…or a lot more in this one:

How I wish Photoshop would embrace AI

AI/ML, ControlNet

Years ago Adobe experimented with a real-time prototype of Photoshop’s Landscape Mixer Neural Filter, and the resulting responsiveness made one feel like a deity—fluidly changing summer to winter & back again. I was reminded of using Google Earth VR, where grabbing & dragging th

Nothing came of it, but in the time since then, realtime diffusion rendering (see amazing examples from Krea & others) and image-to-image restyling have opened some amazing new doors. I wish I could attach filters to any layer in Photoshop (text, 3D, shape, image) and have it reinterpreted like this:

Magic Insert promises stylistically harmonized compositing

AI/ML

New tech from my old Google teammates makes some exciting claims:

Using Magic Insert we are, for the first time, able to drag-and-drop a subject from an image with an arbitrary style onto another target image with a vastly different style and achieve a style-aware and realistic insertion of the subject into the target image.

Of course, much of the challenge here—where art meets science—is around identity preservation: to what extent can & should the output resemble the input? Here it’s subject to some interpretation. In other applications one wants an exact copy of a given person or thing, but optionally transformed in just certain ways (e.g. pose & lighting).

When we launched Firefly last year, we showed off some of Adobe’s then-new ObjectStitch tech for making realistic composites. It didn’t ship while I was there due to challenges around identity preservation. As far as I know those challenges remain only partially solved, so I’ll continue holding out hope—as I have for probably 30 years now!—for future tech breakthroughs that get us all the way across that line.

Day & Night, Magnific + Luma Edition

AI/ML, Relighting

Check out this striking application of AI-powered relighting: a single rendering is deeply & realistically transformed via one AI tool, and the results are then animated & extended by another.

Meanwhile Krea has just jumped into the game with similar-looking relighting tech. I’m off to check it out!