iiinova — where your digital self keeps evolving

Paul

@paulpw.bsky.social

Reposted by Paul

VisionBernie - Bernhard Egger @visionbernie.bsky.social · 22/01/2026

dAIper-Hack A hackathon to turn ideas into impact - hands-on innovation event jointly organized by @fau.de and P&G - three days - modern and classical Computer Vision and AI approaches - at the Scientific Center for AI and SuperTech Apply now! daiper.tf.fau.eu (latest by February 6th)

024

Paul @paulpw.bsky.social · 22/01/2026

🙌 Huge thanks to my co-authors: @jadgardner.bsky.social, @andreead-a.bsky.social, William Smith and @visionbernie.bsky.social (5/5)

000

Paul @paulpw.bsky.social · 22/01/2026

VENI improves scalability compared to RENI++ and has a well-behaved, unique latent space, making it useful in downstream tasks such as inverse rendering, relighting and other vision problems that rely on accurate illumination priors. (4/5)

100

Paul @paulpw.bsky.social · 22/01/2026

To achieve rotation equivariance in the encoder, we use a novel Vector Neuron Vision Transformer, that combines a Vector Neuron transformer with the image patch projection of Vision Transformers. In the decoder, rotation equivariance is achieved by using a rotation equivariant neural field. (3/5)

The Model architecture we use, VN-ViT encoder and RENI++ decoder

100

Paul @paulpw.bsky.social · 22/01/2026

We introduce VENI, a rotation equivariant variational autoencoder designed as a natural illumination prior. Compared to the current state-of-the-art illumination prior (RENI++), VENI achieves higher reconstruction quality and a more well-behaved, unique latent space. (2/5)

110

Paul @paulpw.bsky.social · 22/01/2026

Excited to share our Paper - VENI: Variational Encoder for Natural Illumination 🌐 🔗 Project page: paul-pw.github.io/veni/ 📄 Paper: arxiv.org/abs/2601.14079 👩‍💻 Code: github.com/paul-pw/veni 🧵 (1/5)

Image comparing the uniqueness of VENI to the Uniqueness of RENI++ by showing an interpolation between two optimized images and showing the "Reconstruction Cosistency": Image error vs Latent error of the two models.

132