Move the light on any picture or clip

Every still in this article comes from one character. She is a Plnty picture too: a character sheet, a close-up beside a full-length shot on a white background, made with image-gen on GPT Image 2 in two plain-language edits. The first pushed an earlier portrait into an avant-garde look, with sculptural lips, long chrome nails and geometric metallic hair, and asked it to keep her identity, pose, outfit and the split layout. The second read "change the hair to pink with spikes and metal elements". A sheet like this gives every later step the face and the whole outfit to work from.

Time Machine found the scenes. It takes a picture and a year and returns up to six pictures as that year might have made them, each a different kind of picture and a different scene. Given the sheet and 1927, it painted her among calla lilies, put her in a wig workshop and on a wet street, and set her at the head of a line of dancers under a gilded dome. A second run, at 1970, gave a dressing room mirror among its six. The portrait, the dancers and the dressing room are the scenes relit below.

A picture has no geometry. To move its light you need to know which way each surface faces, the cheek turned toward the lamp and the shoulder turned away, and how far each part stands from the camera. Relight gets both from Depth Anything V2, an open depth model, which it runs inside your browser on your own graphics card through WebGPU, with WebAssembly where WebGPU is not available. The model reads the picture once and returns a depth map, refined against the picture so its edges keep to the subject's outline. Relight turns the map's slopes into the direction every pixel faces and keeps the depth itself for shadows.

From there each drag is a single shading pass in WebGL, on your machine. A soft light wraps further round a form, nearer shapes cast shadows across whatever stands behind them, and a light set behind the subject draws a rim on the edges that face it. The depth also tells Relight what is near and what is far, so the background can keep its own light while the subject takes the new one. Nothing is uploaded and nothing is charged while you drag. You can try a light from every side of the frame in less time than one render takes, and spend credits only on the light you keep.

Why the preview goes with the render

The preview has a second reader. Words describe light badly: "from the right" can mean the subject's right or the frame's, and in our tests one model read it as the subject's. So Run sends the preview along with the picture, described to the model as a rough sketch of the lighting plan, and names each light against the edges of the frame. In Plnty's relight tests, five editing models put the light on the asked side in 30 of 30 renders with the preview attached, against 23 of 30 from words alone.

Light it like a set

Start with one light and drag it slowly from one edge of the frame to the other, watching which side of the face falls into shadow. Stop where the shadow shapes the face the way you want, then add a second light behind the subject and move it until a rim separates the figure from the background. A third light is for colour, or for a source inside the scene such as a lamp or a stage light. Keep the fill low when the shadows should read, and set the lights against each other: warm against cool, or a hard key against a soft rim.

The preview is a plan for the light, rough on purpose. It shows where each light comes from, its colour and which side falls into shadow, and leaves everything else to the render. Run sends the model three things: your picture, the preview, and one sentence for each light that names where it sits in the frame. The model then lights the original picture itself, with the shadows, highlights and reflections the preview only sketches, and the picture you started from stays on the board as it was.

Two models for a picture

We tested before we chose: nine models over two rounds, 126 renders, every model in a round given the same photos and lights. The models built only for relighting came last. One rebuilt the faces and backdrops it was given, and another takes four fixed directions, so it cannot follow a light you drag. General image editors given the preview did best, and the card renders on two of them.

GPT Image 2.5 is the default and the quicker of the two. It stays close to the picture you started with, and it tends to hold a light back, so a soft or faint light can come out quieter than the preview. Nano Banana Pro leans the other way: it commits to the light, and it is the more dramatic of the two. On the dancers it turned an even stage wash into a red key and a gold backlight, with haze hanging between them. Switch to it on the model pill when a scene should change its mood.