iiinova — where your digital self keeps evolving

Alistair Foggin

@alistairfoggin.bsky.social

PhD Researcher in Computer Vision, Graphics and Learning. Exploring image relighting and temporally-varying 3D environment reconstruction from sparse data. University of York alistairfoggin.com

Posts Replies Media

Alistair Foggin @alistairfoggin.bsky.social · 23/04/2026

Feel free to get in contact if you are interested, and you can see the paper online from the project page: alistairfoggin.com/projects/cro...

alistairfoggin.com

CroCoDiLight: Repurposing Cross-View Completion Encoders for Relighting

Disentangle CroCo latents into lighting and scene intrinsics, edit lighting for shadow removal, albedo estimation, relighting and lighting interpolation.

000

Alistair Foggin @alistairfoggin.bsky.social · 23/04/2026

Very excited to be in Rio for #ICLR2026 presenting my poster for CroCoDiLight! If anyone is also here, please pop by and say hi! I'd love to chat. I'll be at poster session 2 in pavilion 4 this afternoon

151

Alistair Foggin @alistairfoggin.bsky.social · 27/02/2026

That sums up CroCoDiLight: Repurposing Cross-View Completion Encoders for Relighting! (Cross-view Completion for Disentangling Lighting) Project Page: alistairfoggin.com/projects/cro... Paper: openreview.net/pdf?id=GKvb3... [7/7]

alistairfoggin.com

CroCoDiLight: Repurposing Cross-View Completion Encoders for Relighting

Disentangle CroCo latents into lighting and scene intrinsics, edit lighting for shadow removal, albedo estimation, relighting and lighting interpolation.

050

Alistair Foggin @alistairfoggin.bsky.social · 27/02/2026

Demo 3: Albedo estimation Similar to shadow removal, working within lighting latent space, we can estimate the scene albedo with lighting removed. [6/7]

On the left there are two images with lighting. One is a stone hallway with arches and bright light shining through the windows between pillars. The other is a bunch of pots and pans over a stove on the ground with light reflections. On the right are the processed versions of these images with the lighting all removed, leaving just the base colour of the scene and objects. These are the estimated albedo images of the scenes.

261

Alistair Foggin @alistairfoggin.bsky.social · 27/02/2026

Demo 2: Shadow removal A model component that works in lighting latent space to map the shadowed latent to a shadow-free latent, effectively removing the shadows. [5/7]

On the left, there are two images with shadows. One is of a toy rollercoaster, with intersecting shadows cast from an external light. The other is of grass and pavement with a person casting a shadow using a branch. On the right are processed versions of the images with all the shadows removed using the model.

150

Alistair Foggin @alistairfoggin.bsky.social · 27/02/2026

Demo 1: Lighting stabilisation of time-lapses We can freeze the lighting so that the only changing aspect is the underlying structure of the scene! [4/7]

261

Alistair Foggin @alistairfoggin.bsky.social · 27/02/2026

Our method disentangles lighting from intrinsics. It takes pixel-aligned image pairs and swaps the lighting. The swapping is forced through a bottleneck so that the intrinsic patches retain image structure. [3/7]

A diagram of two input pixel-aligned images of a building and parking lot in morning and light. Then there are arrows indicating the swapping of the lighting and keeping of intrinsics. These lead into new relit images where the lighting is transferred between the image pair.

140

Alistair Foggin @alistairfoggin.bsky.social · 27/02/2026

We reason that the CroCo encoder from @naverlabseurope.bsky.social embeds lighting information in latent space. In training it must implicitly extract lighting, geometrically cross-project, and then relight the patch. [2/7]

A diagram showcasing how cross-view completion must reason about images. To fill in a missing patch, the model must estimate lighting from other visible patches in the main image and from the paired image, remove the lighting from the paired image, geometrically cross-project it, and relight it to match.

130

Alistair Foggin @alistairfoggin.bsky.social · 27/02/2026

Very excited to announce my first paper, CroCoDiLight! Co-authored with @willsmithvision.bsky.social and accepted at #ICLR2026. Image relighting, shadow removal, and albedo estimation, all in our single base model with swappable task components. See you in Rio @iclr-conf.bsky.social! [1/7]

Teaser image for CroCoDiLight paper. Three panels showing the same scene of a street in three different lighting conditions: sunny with shadows, even lighting without shadows, and night time with lights switched on. A badge for ICLR 2026 is in the bottom right, and CroCoDiLight is in big letters across the top.

191