✦
Summary
Iris-Z is Speridlabs' general image generator that produces appearance and geometry together in one model. One set of weights does four tasks: prompt to image, prompt to image plus depth, image to depth, and depth to image. It is a single diffusion transformer with one backbone and no separate depth head or second network, in two sizes: Iris-Z-big at 32.2B parameters and Iris-Z-small at 9.1B, at 512 and 1024 resolution. The weights and SDK are free under Apache 2.0.
Key Features
Prompt-to-image generation
Prompt-to-image-plus-depth (RGBD) in one pass
Image-to-depth estimation from a photo
Depth-map-to-image generation
Single backbone with one set of weights, no depth head and no second network
Two sizes: Iris-Z-big at 32.2B and Iris-Z-small at 9.1B parameters, at 512 and 1024 resolution
Beats specialist depth models (MoGe-3) on standard depth benchmarks
Python API and CLI with seed, resolution, steps and guidance controls
About Iris-Z
The model is the first general-purpose generator reported to beat state-of-the-art specialist geometry models, outperforming MoGe-3 on standard depth benchmarks. Install it with git clone of the iris-z-sdk repo, weights auto-download from Hugging Face, and use the CLI iris_z.sample or the Python IrisZPipeline API with controls for seed, resolution, steps and guidance. A Hugging Face Space demo is linked from the model card, and the project page is speridlabs.com/research/iris-z.
Note that Iris-Z also reads the gated FLUX.2 autoencoder, which requires accepting Black Forest Labs' terms. Iris-Z is aimed at developers and researchers who need image generation and depth estimation together, for example for 3D-aware editing, relighting and scene understanding, without running two separate models.