Speridlabs has published Idol, a model that turns a single image into a three-dimensional Gaussian splat. Its research page carries October 28, 2026 publication metadata. Code, weights and a demonstration are available. The representation can be viewed from different directions, including predicted portions absent from the input.
What has actually been released
The model card specifies .ply output, 32,768–1,048,576 Gaussians and approximately 7.2 GB peak GPU memory. The core model has about 0.66 billion parameters, but the complete pipeline includes additional components. The core size therefore does not describe every execution requirement.
The SDK is labelled Apache 2.0. Its DINOv3 and FLUX.2 dependencies require separate access and acceptance of their terms. The code also saves the processed input image and predicted camera. An open licence for one component does not cancel the conditions attached to others.
Speed needs a defined configuration
The introduction includes an “under two seconds” claim. However, the published table for 50 steps and 262,144 Gaussians lists 3.4 seconds on B300, 19.1 on RTX 4090 and 46.2 on RTX 3080. These are author measurements, medians across eight images, not SIExpo tests. The source does not reconcile its introductory wording with those conditions; we do not treat it as a general guarantee.
That discrepancy has practical consequences. Waiting may be acceptable for an artist creating one object. In an interactive workflow with many attempts, each regeneration changes the working rhythm. Neither application can be evaluated from the shortest promotional number alone.
Predicted shape is not measured geometry
Our editorial interpretation is that a single input view supplies limited evidence about hidden surfaces. A generated back can look convincing while differing from the real object. Increasing the number of elements in the representation does not turn that hypothesis into a measurement.
For artistic ideation, this may be useful: an author receives a proposal to inspect and modify. For a product catalogue requiring geometric accuracy, we would request a separate validation step. We do not claim that the output is a finished engineering model, production mesh or verified replacement for scanning.
How we would evaluate it before adoption
A practical trial set could contain objects with distinct faces, thin features, handles and occluded surfaces. We would preserve photographs of their real rear surfaces solely for later comparison. This could distinguish recovery of a known shape from generation of a plausible alternative.
We would separately measure first and subsequent executions, preparation and retry time, and compatibility between the .ply output and the intended viewer. This is a proposed evaluation, not an experiment we conducted. Its purpose is to assess the complete local workflow.
An evaluation should also record unsuccessful attempts rather than display only attractive examples. An artist needs to know how much correction an ordinary result requires, and whether the same input remains useful at the chosen level of detail. Those questions concern practical suitability rather than a universal model ranking.
Idol is noteworthy as an accessible spatial-generation release with actual published artifacts. Its limitations belong in the news too: dependency permissions, hardware-dependent latency and no guarantee that unseen parts have been reconstructed faithfully.
*Cover: AI illustration; a conceptual Gaussian object, not output from an Idol test.*
*Editorial verification: 30.10.2026, 19:35:25 +0200 (Europe/Sofia).*
