Summary
According to a post on X by @speridlabs, Sperid Labs has open-sourced Iris-Z, a model that generates images and depth from one backbone. The model can estimate depth or condition on it and comes in 9B and 32B sizes, with weights, code and a demo released under the Apache 2.0 license. Sperid says its 32B model beats task-specific depth models such as MoGe-3 on Marigold's benchmarks, with a mean AbsRel of 2.92 against 3.12. These figures are the company's own claims as reported in the post.
AI-assisted summary based on the listed source.
What happened
Iris-Z generates images and depth from one backbone and can estimate or condition on depth, coming in 9B and 32B sizes with weights, code and a demo under Apache 2.0. Sperid says the 32B model beats task-specific depth models such as MoGe-3 on Marigold's benchmarks (2.92 vs 3.12 mean AbsRel). ...
Why it matters
Developers and researchers working on depth estimation or image generation can inspect the weights and code under a permissive license. Because the benchmark results come from Sperid itself, independent evaluations on the same Marigold benchmarks would be the next thing to watch. The choice between 9B and 32B sizes may also matter for teams weighing compute costs against accuracy.
Signal Intelligence
Signal Strength 91%
Technical label WATCH
Public Interest 46
Category OPEN SOURCE
Reader Depth TECHNICAL
Signal Strength reflects source quality, relevance, freshness and evidence. Public Interest helps organize discovery; it is not proof of truth.
Public Interest components
Recognizable Entity Score 0
Practical Impact Score 56
Novelty Interest Score 72
Consequence Score 46
Curiosity Score 84
Shareability Score 51
Why this is here
VQV surfaced this signal because it is recent, relevant to Robotics, connected to Scobleizer News Pulse (X).
