Two years of generative art direction for the Libera Awards. Ben Golden

Before

The first viable concept became the final concept

Through 2024 the package was 2D, made by hand. Every visual decision had to work across the whole show from day one, because there was no room left to explore once a direction locked in. Forty-plus categories. Two hundred-plus album covers in every palette you can imagine, each one to accommodate rather than fight. Names of wildly different lengths. A premium bar, on a fixed budget and timeline.

Under those constraints, the first concept that technically worked was the one you shipped. Trying a second one would've cost the show.

2024 nominee screen, Best Alternative Rock Record: flat 2D artwork built by hand — **2024 versus 2026**The same category, Best Alternative Rock Record. Flat 2D built by hand, then a 3D environment.

2026 nominee screen, same category, set in a generated 3D environment — **2024 versus 2026**The same category, Best Alternative Rock Record. Flat 2D built by hand, then a 3D environment.

2025

Changing how the images got made

Volume of options is the thing generative tools are actually good at. Everything downstream of the options is judgment.

Concepts at volume. I wrote around fifty prompts for candidate environments and built a ComfyUI pipeline that could render hundreds of frames overnight. In prior years I'd have had time to sketch maybe three directions. Now one more option cost next to nothing, so the job became about choosing well.

From image to space. The strongest frames went through depth estimation and turned into 3D meshes with real spatial depth. A generated hallway stopped being a backdrop and became an actual set, something a camera could move through, light, and hang a nominee's artwork inside.

Textures. An experimental ComfyUI-to-Blender integration generated bespoke textures instead of pulling from a stock library, so the surfaces actually matched the concept frames they came from.

The result was a full 3D package, delivered on the same schedule the flat version used to take.

Eight generated 2025 category backgrounds: concrete corridors, arcades and stepped chambers, each in a different marbled palette — **2025 category backgrounds**Eight of the set. Generated, art-directed, and still flat. One more option cost almost nothing, so the work moved to choosing.

Blender with a generated environment loaded as geometry: marbled walls, an ornate frame playing a nominee video, statues either side, lights and a camera rigged on it, and a shader node graph open below — **The mesh, dressed as a set**Once the frame is geometry it takes lights, a camera and a nominee's artwork on the wall. What comes back from depth estimation is rough, and that is fine: it only has to hold a camera move and take light.

2026

Feeding the agents

The 2026 show extended the same practice: generated stills, video, 3D objects and environments, all art-directed the same way I'd always done it. The two kinds of 3D took different routes. Environments went through depth estimation into mesh; objects were generated directly, with Meshy running inside ComfyUI. Those assets are what the agent pipeline assembled into finished nominee screens. That production is its own case study, and it depended on this one settling the visual language first.

A generated 2026 tunnel interior beside its depth pass, the same geometry rendered as a greyscale distance map — **A 2026 frame and its depth pass**Distance read as brightness: near surfaces light, the mouth of the tunnel black. This frame went the whole way, render to depth to mesh, and ended up as a set the camera moves through.

What stayed human

Taste became the whole job

Art direction. Picking from hundreds of frames. How the generated passes got combined into one coherent package. Every judgment call on whether the result looked like an awards show or a cheap demo.

Generative tools moved the visual ceiling. They didn't move the bar for taste one inch. If anything, taste became the whole job: when one more option costs nothing, choosing badly is the only way left to fail.

There were hard edges too. Nominee artwork is fixed, never generated or altered, and the environments had to hold two hundred covers, not just one. Names of every length had to sit legibly in the same layout. A gorgeous frame is worthless if it can't hold up in the fortieth category as well as the first. The pipeline got built around those constraints, not around whatever the models happened to be best at.