Adobe’s TokenLight lets you move 3D lights in a 2D photo

Commercial photographers know the pain of bad lighting all too well. If your key light is off by an inch on set, fixing it later in post usually means hours of painstaking masking and fake-looking dodging.

Adobe introduced TokenLight, a new relighting research project that turns static 2D images into interactive virtual photo studios.

Virtual lighting with slider controls

Instead of applying flat color overlays, TokenLight treats image relighting as a conditional generation problem.

The system takes core lighting parameters—intensity, color, ambient light, diffuse level, and 3D light coordinates—and encodes them into discrete attribute tokens fed into a Transformer model.

Creators can use simple directional controls to drop a virtual lamp anywhere in a photo. You can push the light deeper into the background or pull it closer to a foreground subject, and the scene will immediately recalculate shadows and highlights across every surface.

The diffusion controls work like a continuous slider. Dial it down for razor-sharp, hard shadows; dial it up for soft, diffused morning glow.

Learning physics without an explicit 3D engine

What makes TokenLight stand out is how it handles scene understanding.

Traditionally, realistic relighting required explicit inverse rendering—estimating depth maps, surface normals, and material albedo before calculating how light bounces. TokenLight skips all that.

Adobe trained the model on a massive synthetic dataset paired with lighting annotations, supplemented by a small set of real captures. Remarkably, without any explicit 3D inverse rendering supervision, the model figured out how light interacts with geometry, occlusion, and materials on its own.

It can even handle tricky cases like placing lights inside translucent objects or casting believable refractions through transparent glass.

It is not flawless yet. Reflections in mirrors still do not update dynamically when you move the light source. And since Adobe has only published the technical paper so far, do not hold your breath for open-source weights.

Why it matters

For product designers, digital artists, and photographers, this points to a massive upgrade for creative workflows. Being able to change the time of day, add rim lighting, or tweak ambient fill in a finished 2D photo removes a huge chunk of production anxiety.

For the AI industry, TokenLight demonstrates that neural networks are steadily absorbing tasks once reserved for dedicated 3D renderers. By turning physical lighting coordinates into language-like tokens, Adobe has found a natural bridge between generative models and practical user interfaces.

In our view, Adobe's attribute-token approach is brilliant because it gives creators continuous, fine-grained control instead of unpredictable text prompts. The catch is delivery: Adobe rarely open-sources this kind of core tech, meaning creators will likely have to wait until it trickles into a future paid Creative Cloud release.