Sperid Labs debuts Iris-Z joint image-depth model

Unifying geometry and appearance into a single token sequence is the architectural shift generative 3D has been waiting for. Instead of chaining fragile multi-model pipelines or bolting post-hoc depth estimators onto 2D diffusion, Iris-Z proves that general-purpose foundation models can internalize spatial structure and surpass bespoke depth specialists.

• –Single-backbone multimodal tokens: Storing depth as normalized log ray distance and passing it through the shared autoencoder allows image pixels and geometry to condition each other iteratively across every denoising step.
• –Four workflows in one checkpoint: A single model natively handles text-to-image, RGBD generation, monocular depth estimation, and depth-to-image re-rendering without switching weights or maintaining separate ControlNet adapters.
• –Outperforming specialist baselines: Iris-Z-big (32B) edges out MoGe-3 on standard Marigold benchmarks (2.92 vs 3.12 AbsRel), proving that broad generative pretraining can beat task-dedicated vision architectures.
• –High compute overhead: Running 40 diffusion passes on a 32B model takes 15 seconds on an H200 compared to MoGe-3’s 0.4 seconds, making Iris-Z currently better suited for high-fidelity asset generation than real-time robotics or spatial computing.