Iris-Z, the new generative model from Speridlabs, does something traditional image generators do not: it creates the image and the depth map of a scene at the same time, with the same weights. According to the lab, it is the first general-purpose generator to beat specialist depth models such as MoGe-3 on standard geometry benchmarks.
Presented in November 2026, with a paper on arXiv and weights published on Hugging Face, the model was developed by Alejandro Cañada and Chema Garabito of Speridlabs and builds on FLUX.2 by Black Forest Labs.
One model, four tasks
With a single set of weights, Iris-Z performs four tasks: generating an image from a prompt, generating the image together with its depth map, estimating the depth of a photo, and creating an image from a depth map. There is no depth head and no second network: depth is stored as normalised log ray distance, encoded as an image by the same autoencoder and placed in the same token sequence, so image and depth condition each other at every denoising step.
Two versions: big and small
Iris-Z comes in two variants. The larger one, Iris-Z-big, has 32.2 billion parameters; the smaller, Iris-Z-small, has 9.1 billion. Both generate at 512 and 1024 pixel resolutions. According to the official model card, the big version needs about 112 GB of GPU memory for all tasks, and the small one about 33 GB — in other words, a heavy model aimed at workstations with high-end GPUs, not casual use on laptops.
What the benchmarks show
On depth estimation benchmarks, Iris-Z-big beat MoGe-3 under both compared protocols. Under the MoGe-3 protocol, the mean relative error (AbsRel × 100, lower is better) was 3.56 for Iris-Z-big versus 3.70 for MoGe-3; under the Marigold protocol, it was 2.92 versus 3.12. The small version also beat MoGe-3 under the Marigold protocol (3.11 versus 3.12), but fell behind under the MoGe-3 protocol (3.84 versus 3.70).
There is a cost, though: the specialist remains far faster and cheaper per image. In the benchmark setup, one Iris-Z-big prediction takes 15 seconds on an H200 GPU, versus 0.4 seconds for MoGe-3. For the standalone task of estimating the depth of a photo, with a single pass, Iris-Z-big takes 1.44 seconds at 512 pixels — much more practical for everyday use.
Why generate image and depth together
When a model learns appearance and geometry at the same time, the generated image is born with a coherent sense of space: what is near and what is far, and on which plane each object sits. That is useful for depth-guided image editing, for creating 3D scenes from text, for robotics and for computer vision — areas where knowing the distance of each pixel matters as much as its colour.
Open source and access
The Iris-Z code is released under the Apache 2.0 licence, and the weights are on Hugging Face. To run the model, you must accept the terms of the FLUX.2 autoencoder, which is gated, and use an Nvidia GPU with driver 580 or newer, with CUDA 13.0. The official SDK provides commands for all four tasks, including generating image and depth together from a prompt such as “a lighthouse at dusk”.
Read also: FLUX 3 Image Klein: Black Forest Labs releases 5.6B open-weights model for image generation and editing (in Portuguese)
Sources
- Speridlabs — Iris-Z (official project page)
- Hugging Face — speridlabs/iris-z (weights and model card)
- arXiv:2611.16010 — Iris-Z: Scaling One Generative Model Past Specialist Geometry Models
- GitHub — speridlabs/iris-z-sdk (code, Apache 2.0)
Featured image: official Iris-Z example — image and depth map generated from the same prompt. Credit: Speridlabs / Hugging Face.
This article was produced with the assistance of artificial intelligence, cross-checked against public posts and internet sources. Check the original sources to confirm the information before sharing.
Por Muse
