Decades in the making: How Adobe Research brought Relight to Photoshop

For decades, researchers have chipped away at the problem of realistically editing lighting across an entire image. Now, Adobe Research and the Photoshop team have brought that capability to creators with Relight, just a year after its debut as a 2025 MAX Sneak. Relight’s arrival in Photoshop reflects both years of scientific progress and a close collaboration between researchers and product teams who share a commitment to expanding creative possibilities without compromising users’ control over their work.

Most photo edits are local, but light affects an entire scene, making it particularly difficult to manipulate.

“Light is a global effect,” says Zhixin Shu, the Adobe Research Scientist behind Relight’s model development. “You change lighting, you change a lot of other things.”

Realistic relighting requires understanding how light interacts with objects, shapes, materials, and positions. The challenge is especially apparent in portraits.

“Our visual perception is very tuned to what humans look like,” Kalyan Sunkavalli, the Principal Scientist leading the team, explains. “If there are small errors, it immediately jumps out at us.”

That’s why relighting has been an open challenge in the world of computer graphics for 30 years, and why Adobe Researchers have been innovating in the space all this time.

The research has evolved from early physics-based methods to deep learning approaches. During his 2016 internship, Shu worked on a paper that introduced a method for transferring lighting from one portrait to another while accounting for facial geometry. By considering the shape and contours of a person’s face, the technique could produce more realistic lighting adjustments across different portraits.

“At the time, we were trying to give users control over lighting,” Sunkavalli recalls of those earlier efforts. “But the quality of the images you would get, and the precision of the control, was not even close to what we have now.”

A long-standing problem meets new breakthroughs

Modern image models have become increasingly capable of producing realistic images, learning patterns in how light interacts with objects and surfaces. For Adobe Research, advances in generative AI created even more opportunities.

“Now you have this tool that’s really good at producing images,” Sunkavalli says. “The problem becomes, ‘How do you steer this tool to do the precise tasks that we want to do?’”

From the earliest stages of Relight’s development, creative control was a central consideration– more specifically, developing AI that responds to creators’ intentions rather than making the decisions for them.

The second development was data, which became a major focus of Shu’s research. He created a pipeline for high-quality synthetic images that could represent the same scene under different lighting conditions. By training on these examples, the team could teach generative models to manipulate illumination without altering the rest of an image.

Bringing Relight beyond the MAX stage

At Adobe MAX 2025, Shu unveiled Project LightTouch, a Sneak showcasing the team’s research and offering audiences an early glimpse of how AI could transform lighting in existing photographs.

Photoshop’s Director of Product Management Amy Durocher recalls being immediately drawn to the technology when she first saw the research ahead of its MAX debut: “I want this feature and I want it really bad.”

Having recently helped bring another Adobe Research innovation, Rotate Object, into Photoshop, her enthusiasm was rooted in an established working relationship with Sunkavalli. Their ongoing exchange helped Durocher identify promising technologies early and explore how they might translate into useful creative tools.

Durocher began exploring how the research could translate into an efficient editing experience. The biggest question was whether the model could become fast enough to support the kind of interactive experience Photoshop users expect. Early versions took approximately nine seconds to generate a preview after a lighting adjustment, a delay that Durocher knew would need to come down significantly for the feature to be viable.

Sunkavalli was confident the team could solve it.

Putting creators in control of the user experience

Rather than developing the model and user experience separately, Adobe Research and the Photoshop team worked in parallel. Photoshop Product Designer Dana Jefferson helped shape the interface, while early prototypes allowed the teams to gather user feedback before engineering integration.

“You don’t just train a model and then throw it across the wall and then build a user experience,” Sunkavalli explains. “The user experience dictates what the model should look like.”

The teams wanted to ensure that the experience gave creators the freedom to make their own lighting decisions, with AI supporting the editing process rather than determining the outcome.

That philosophy is reflected in how Relight works in Photoshop. Creators can place virtual light sources within an existing photograph, reposition them, and adjust their lighting properties to experiment with how a scene is illuminated. Instead of relying on a text prompt to describe the desired result, they can make visual adjustments themselves and refine the effect until it matches their creative vision.

For Durocher, who spent a decade as a designer before moving into product management, preserving that level of control was particularly important. She understands firsthand that creative work involves making deliberate decisions, often through small adjustments that bring an image closer to what its creator envisioned.

“I want to make the choices. I want to be in the driver’s seat,” she says.

That commitment to creative control also shaped the team’s response to prerelease feedback. The teams had developed a two-stage workflow that let creators experiment with lighting through faster, lower-resolution previews before generating a higher-quality final image, effectively tackling the speed issue.

However, users identified subtle differences between the lighting they adjusted in the preview and the final, full-resolution image. Although the changes were relatively minor, the feedback showed how important consistency was to creators who had carefully composed their edits.

“Users really care,” Shu explains. “They spend time authoring it. They don’t want anything changed.”

The researchers responded by refining the model through a technique called distillation, developing a one-step version that kept lighting consistent between preview and final output.

Durocher says the team’s willingness to engage directly with user feedback was particularly meaningful. Beyond responding to her prioritized feedback, the researchers read user forums and watched creator videos themselves to better understand how people worked with the feature.

“For researchers to engage that way is very special,” she says.

Bringing decades of research into creators’ hands

Only a year after its MAX Sneaks debut, Relight launched in Photoshop. For the team, the launch is an opportunity to see how creators will work with the technology and learn what still needs improvement.

Relight’s journey also demonstrates the importance of maintaining close relationships between researchers and product teams.

“We’ve shipped so many features together, and that doesn’t happen out of nowhere,” Durocher says. “It takes work. I think it’s a great model for other research teams to follow.”

More broadly, Relight illustrates how a shared focus on creative control can bring research and product development closer together. From early model development to interface design and user testing, the teams continually shaped one another’s work to ensure the technology served creators’ needs.

Ready to try Relight? Open the latest version of Photoshop, select an image layer, and navigate to Layer > New Relight Layer to start experimenting with lighting. Learn more about Relight and how to get started.

Adobe Research, Photoshop, ART, and NextCam worked together to bring Rotate Object to life. Key contributors include:

• Adobe Research: 
  • Core contributors: Zhixin Shu, Matthieu Gallard, Yiqun Mei, Kalyan Sunkavalli. 
  • Other contributors: Vojtech Krs, Yannick Hold-Geoffroy, Xin Sun, Handong Zhao, Dipti Thakkar, Nathan Carr 
• ART Sensei/Services: Joel Baer, Albright Djokoto, Amy Durocher, Betty Leong, Dongmei Li, Mark Maguire, Rick Mandia, Hannah Nicollet, John Peterson, Katie Riley, Yukie Takahashi, Raymond Yan, Zhengyun Zheng, Abhijit Apte, Shayan Chandrashekar, Noah Hoffman, Katia Kliuyeva, Rahul Koul, Nithesh Gangardhar Salian, John Tran, and Sangeeta Varma
• Photoshop, Prototype Design & Adobe Design Research: Michael Cragg, Sapna Gathani, Dana Jefferson, Quynh Nguyen, and Evan Shimizu
• NextCam: Mingming He and Marc Levoy
• Special thanks to: Daniel Presedo, Cory Shubert, and Claudia Rodriguez

Wondering what else is happening in Adobe Research? Check out our latest news here.