Open-source Iris-Z unifies image generation and depth in one backbone, beating MoGe-3
victormustar · x · 2026-11-12
Speridlabs released Iris-Z, billed as the first general image model to combine appearance and geometry in a single set of weights — no depth head, no second network.
- Built on a pretrained text-to-image diffusion transformer: depth is stored as normalized 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
- Available in 9B and 32B; weights, code and demo are fully open under Apache 2.0
- Results: the 32B model beats specialist depth model MoGe-3 on its own benchmarks (mean AbsRel 3.56 vs 3.70; 2.92 vs 3.12 on Marigold's sets), though it needs 15s per 512² image versus MoGe-3's 0.4s
More from Models
- VulnCheck audit: only 0.57% of Anthropic Glasswing's 37,966 findings ever fixed — WeldPond · 2026-11-12
- On PostTrainBench, AI post-training now beats the human-tuned SmolLM3-3B from the same base — maksym_andr · 2026-11-12
- Claude flags mispronunciations in video audio it claims it can't hear — RileyRalmuto · 2026-11-12
- Heavy Users Say Claude's Safety Rails Block Even Basic Chemistry Work, Forcing Fallback to Opus — rbhar90 · 2026-11-12
- Community teams release wave of abliterated Qwen models for 8–24GB local GPUs — udmrzn · 2026-11-12
- Redditor claims Qwen 4 Max slots between DeepSeek v4.1 and Kimi 3.1 Pro in output quality — power97992 · 2026-11-12