Alvin Ghouas

Alvin Ghouas is a software engineer and UX designer whose AI work spans creative interfaces, controllable image generation, and privacy-sensitive machine-learning training. In April 2023, he presented ProductStudio, a side project described as a 3D editor with Stable Diffusion Img2Depth, at the first AI | UX meetup at Notion HQ in San Francisco. He also publishes under the name Alvin Alexander Ghouas.[2][1][3][4][5]

Ghouas brought ProductStudio to a meetup exploring interfaces built on generative AI. The Latent Space recap, published on 26 April 2023, identifies him as its presenter and describes the demonstration as a “3D editor with Stable Diffusion Img2Depth.” In his own post that day, he called ProductStudio “a side project I've been working on.”[1][3]

In that April 2023 post, Ghouas argued that maturing models created an opportunity to build “user-friendly and safe tools on top of them.” He saw user-centered design as consequential both commercially and for aligning models with human goals and ethical frameworks. This was his argument for the role of interface design in AI, rather than a demonstrated alignment result.[3]

Ghouas coauthored Controlled and Conditional Text to Image Generation with Diffusion Prior, whose 1 August 2023 version lists an Adobe Applied Research affiliation. His collaborators included Pranav Aggarwal, Hareesh Ravi, Naveen Marri, Baldo Faieta, Ajinkya Kale, and other coauthors. Their work examined how to constrain generated images to useful visual domains and control their color palettes.[4]

The team's architecture separates generation into two stages: a diffusion prior produces an intermediate CLIP image embedding from text, and a diffusion decoder turns that embedding into an image. After training a shared decoder, the researchers trained separate priors for textures, rasterized vectors, and isolated objects while keeping that decoder intact. The domain priors differed in their training data; the vector prior produced raster images suitable for subsequent vectorization, rather than directly generating SVG files.[4]

For palette control, the researchers added a color-histogram token to the prior's inputs. The paper illustrates how a berry prompt could produce blueberries under a blue exemplar palette, compared with raspberries without color conditioning. The authors also reported a material limitation: the histogram generally took priority over color words in the prompt unless the colors overlapped, and illustration exemplars could influence the output toward an illustration style.[4]

Before the 2023 interface demonstration and research publication, Ghouas was named among the inventors on an Adobe patent application filed on 30 March 2021. The invention addressed training on user-generated assets that human developers may be prohibited from viewing directly. His fellow inventors included William Marino, Tim Converse, Shabnam Ghadar, and five others. The application was published on 6 October 2022; the granted patent, US12153705B2, was published on 26 November 2024.[5]

The disclosed system reviews training code for prohibited logging or display of private data, records a hash of approved code, and verifies that code before allowing it to run in a protected environment. Its access mechanisms exclude assets subject to deletion requests or machine-learning opt-outs, while other embodiments impose retention limits and delete training assets after use. This shared invention supplies earlier context for Ghouas's work on AI infrastructure; the patent describes an architecture, without establishing deployment or his individual responsibility for a particular subsystem.[5]