Speridlabs (Alejandro Cañada, Chema Garabito) · Joint image and depth

Iris-Z

Iris-Z is an image generator from Speridlabs that produces appearance and depth together with one set of weights, with no depth head or second network. It starts from a pretrained text-to-image diffusion transformer, and depth is stored as normalised log ray distance, encoded by the same autoencoder and placed in the same token sequence as the image. On MoGe-3's benchmarks the 32B Iris-Z-big reaches a mean AbsRel x 100 of 3.56 against 3.70 for MoGe-3, and 2.92 against 3.12 on Marigold's sets; the 9B Iris-Z-small scores 3.84 and 3.11. The cost is speed: one prediction at 512² on an H200 takes 0.4 s for MoGe-3, 4.8 s for Iris-Z-small and 15.0 s for Iris-Z-big. The page links code, weights, a demo and an arXiv paper (2611.16010), but does not state a licence.

Open the source ↗

Similar 3D & Spatial releases

More from Speridlabs (Alejandro Cañada, Chema Garabito)

  • Idol

    Speridlabs (Alejandro Cañada, Chema Garabito) · Image-to-3D splats · Week 45, 2026

Published by Quanta AI Labs, an AI agency in India. Browse every week in AI News Weekly.