Gray Crawford

Gray Crawford is a designer and engineer whose work includes generative AI interfaces at Luma AI and instruments for live audiovisual performance. In March 2026, he directed research and training for Latent Blend, a custom in-house model for image embedding interpolation, and contributed core design and engineering to Luma Agents’ agentic canvas.[2]

Crawford’s portfolio records product design work at Luma AI in 2024–25, including model evaluation and finetuning data for Ray2 in January 2025 and core design of Modify Video that June. His 2026 contributions included core design and engineering for Ray3.2 Modify’s multi-keyframe video-to-video feature in June and Luma Scenes’ autoregressive multi-keyframe storyboarding feature in August.[2]

Outside that product work, Crawford developed Neophon, a custom generative music performance instrument using Lyria Realtime, listed in January 2026. It focuses on responsive, high-dimensional interpolation. At Anthropic’s Code with Claude developer conference that May, he performed two three-hour improvisations with the realtime audio generation instrument.[2]

Earlier work provides a detailed example of Crawford’s approach to controlling generative models through movement. In 2019, he collaborated with Aman Tiwari on Xoromancy, a hand-controlled BigGAN image explorer developed during his Carnegie Mellon design thesis.[5][3]

Xoromancy used a Leap Motion sensor to turn hand position and rotation into neural-network inputs. Participants could change generated images through continuous motion, with distinct gesture categories controlling subject and composition. Tiwari was the primary software engineer, using Unity and Python with TensorFlow; Crawford’s thesis describes their collaboration over one month.[4][5]

The implemented mapping exposed fourteen components of BigGAN’s z-vector, seven per hand. Each hand’s movement along three spatial axes controlled three components; its four quaternion rotation components mapped to four more model inputs. Crawford describes independent traversal as a design goal: changing one mapped component should not necessarily change another. Hand rotation still has three degrees of freedom, however, and the four components of a unit quaternion are constrained, so the quaternion-derived inputs are coupled. The mapping therefore exposed seven model inputs per hand without providing seven independent physical controls.[5]

The collaborators tuned the scale of each movement-to-input transform to keep values within a range that produced coherent imagery. Crawford recalls that more extreme values began mixing subjects before dissolving into harsh visual artifacts. Reversing hand movements also reversed the visual changes, allowing users to return to latent-space coordinates. The interface linked the felt position and movement of the hands—proprioception—to image feedback, rather than requiring users to explore one image at a time.[5][4]

Xoromancy first appeared at Carnegie Mellon’s FRAME Gallery on February 22–24, 2019. Crawford observed participants using both broad movements and meticulous adjustments; several reported feeling that they were “mastering” it over time. His thesis also recounts presentations at New York Live Arts’ Live Ideas festival in May and IEEE-GEM at Yale University in June 2019.[3][5]

In his thesis, Developing Embodied Familiarity with Hyperphysical Phenomena, dated May 2019, Crawford argued that unfamiliar interfaces could reward practice rather than merely reproduce familiar controls: “If we constrain our movements to the limited scope of previous media, we might make intuitive artifacts, at the expense of allowing users to develop new intuitions for the new spaces they are gaining access to.” His prototypes primarily used Leap Motion hand tracking in VR to investigate engagement with computed behaviors beyond ordinary physics.[5]

Crawford studied liberal arts at St. John’s College, Santa Fe, during 2011–15, followed by an MA and MDes in Design for Interactions at Carnegie Mellon during 2016–19. In summer 2018, he interned in AR design prototyping at NASA’s Jet Propulsion Laboratory, working on bimanual hand-anchored ProtoSpace interfaces and Leap Motion tracking for HoloLens. From 2019–23, he worked as an interaction design engineer at Ultraleap, prototyping spatial interactions in Tracking and Interaction R&D.[2]

At the 2023 AI | UX meetup, Crawford presented Mapping Image Generation to Movement. According to the April 26 publisher recap, he demonstrated GANBreeder/ArtBreeder mapped to movement with a Leap Motion controller.[1]