PIXLRelight Is a Relighting Demo I Want to Try

PIXLRelight caught my attention because it is trying to do something more specific than the usual “change the mood with a prompt” image edit. It is a relighting system from Miguel Farinha and Ronald Clark at the University of Oxford that focuses on physically controllable single-image relighting. The paper was submitted to arXiv on May 18, 2026.

The interesting part is the control model. The project combines learned image synthesis with intrinsic scene information, then lets the target lighting come either from a reference image or from a path-traced Blender Cycles render. According to the project page and paper, the model can apply arbitrary PBR-style lighting control and run in under a tenth of a second per image.

Why I want to try it

I would like to test it on some of my own photos for three practical reasons: how believable the relighting looks on ordinary images, how much setup the Blender and render-pass workflow actually needs, and whether the extra control is worth the effort compared with simpler relighting tools.

What the public materials show

• The demo page includes interactive examples where the light position changes as you move the cursor.
• The paper describes an inference pipeline that recovers coarse geometry and materials, renders target lighting, and then uses a transformer-based renderer to produce the final image.
• The public GitHub repository includes inference code, sample layouts, and two input modes: target RGB images and Blender Cycles render passes.

If the real-world setup is manageable, this looks like a useful middle ground between casual relighting effects and a full manual 3D relighting workflow.

Links: project page, arXiv paper, GitHub repository.