Nuke 18.0 runs AI workflows right in the node graph

Foundry has released the open beta of Nuke 18.0. The key new feature is a bridge to Griptape, a platform for orchestrating AI workflows: users can run AI workflows right inside Nuke's node graph.

The current build is Nuke 18.0v1 Beta 2. In parallel, Foundry is offering the Nuke 17.2v1 Beta 2 build. The company strongly advises against using beta features in production, and has not yet named a final release date for 18.0.

Pairing Nuke with Griptape

The integration is billed as a bridge between Nuke and Griptape: the AI operation runs not in an external application, but in the same graph where the shot is being assembled. Implementation details — the set of parameters, the supported workflow types and the terms of access to the models — are not spelled out in the beta description.

Gaussian splats in the viewport

The second major part of the beta is addressed to those already working with splats: Nuke got native splat import and rendering in version 17.0 in February, and 17.1 added dynamic splats and relighting following the open beta in July. In 18.0 there are four viewport controls that reduce the cost of dense splat assets: lowering splat density, drawing them as simple colored points, rendering only the splats that fall within the frustum of the selected camera, and turning off spherical harmonics. All four act solely on the 3D viewport and do not change SplatRender's final output. The 3D viewer's color correction can now be switched between sRGB (the default) and off. Foundry does not give the names of the knobs themselves.

A FieldThreshold node has appeared: it builds a density or signed-distance field from thresholds on incoming attributes, and that field can be used to mask other nodes — GeoDeletePoints or GeoGrade, for example. The supported attributes are: position, scale, transparency, color, HSV, distance to the nearest neighbor and distance from the origin. A compositor cleaning up "floaters" can set a threshold on distance to the nearest neighbor and feed the field into GeoDeletePoints, or use GeoGrade to tweak only the low-transparency splats in a scan. Splats can also be selected right in the viewport in vertex selection mode, with FieldShape snapped to the chosen point.

Foundry also lists limitations: the node can be slow on splats scattered unevenly over a large area, and its edits sometimes cause upstream Geo operations to disappear from the viewer. The beta also fixes a SplatRender crash when requesting Deep samples outside the range, a noticeable loss of splat quality when rendering to 2D via SplatRender, a brightness mismatch between the 3D viewer and SplatRender's output, and two frustum clipping errors.

What else is in the beta

Besides splats, the beta adds interaction between Nuke and Griptape, automatic cut detection and H.265 decoding.

How to get into the beta

Builds of the Nuke 18.0 Open Beta are available now to both licensed and trial users.

Why Foundry is embedding AI this way

The pairing with Griptape shows the strategy Foundry has chosen: what is built into Nuke is not a specific generative model, but a gateway to a platform capable of chaining different models into a single manageable workflow. Given the speed at which image and video generators replace one another, a hardcoded model would be obsolete before it reached production.

This approach likely continues the line begun with the CopyCat node: Foundry gives the compositor a manageable tool inside the graph — with versioning, reproducibility and the ability to fix any step. Running AI operations from the node graph looks like an answer to studios' main objection to generative services: the material should preferably not leave the secure perimeter, and every frame must be traceable.

If the trend holds, the boundary between compositing and generation will keep blurring, and demand will likely shift toward specialists who can not only clean a plate by hand, but also assemble a reproducible AI workflow for the shot's needs. While Foundry is not naming a final release date, studios have a chance to test the beta on their own projects and pass on feedback before the stable version ships.

What is behind Foundry's turn toward an external AI platform: pressure from competitors, studio demand for comp automation, savings on in-house development?