Ben Golden

Case studies / DistroKid, for Direct Agents, 2026

Animated spots for DistroKid on a generative stack

Three animated spots shipped for paid social. The stack: open-source image models run locally, image-to-video, and compositing.

  • Client DistroKid, through Direct Agents
  • Role Concept, art direction, generation, compositing, edit
  • Year 2026
  • Tools ComfyUI, open-source image models, image-to-video, After Effects, Premiere Pro

The spots

Three shipped

DistroKid is a music distribution platform. Each spot is animated brand advertising for paid social, which means it's got to hook you in the first second and hold on for the rest.

Out of the Box is the full story. A musician alone in a bedroom studio at night, the upload, and the music leaving the house as light. The whole spot is the moment between finishing a song and it being somewhere other than your own hard drive.

Your Music Everywhere puts the same setup, the desk and the interface and the mic, in places it has no business being. None of the gear changes. Only the world does.

The POV cut of the same idea never leaves the phone. One hand, one screen, and the environments swapping around it: forest, mountain, a street corner, a snowbank. Same argument, made with a camera instead of a set.

Out of the Box
Your Music Everywhere
Your Music EverywhereThe POV cut.

The stack

Generative passes as raw material

The stack is named. The prompts, workflow files, and numbers are not. Knowing the tools gets you about as far as knowing which lenses a DP owns.

Three kinds of pass go into a shot, and none of them is the shot by itself.

Flux with trained LoRAs, run locally in ComfyUI, produces the stills that define a world: its light, its palette, its materials. Running them locally means I can iterate hundreds of times at no per-image cost. Those stills get cleaned and colour-corrected by hand in Photoshop before anything moves.

Image-to-video in Seedance, driven by first and last frames rather than a prompt alone, puts motion into those stills. What comes back is a moving element with the right physics and texture, one that can be placed, timed, and graded just like any other piece of footage.

Compositing is where the shot gets made. Separate generative passes almost never agree with each other on lighting, edge, or motion, so making them read as one image is the same work visual effects has always done. It's the step no model does for you, and the difference between a folder of clips and a spot worth paying for.

One rule organises all of it: deterministic work goes last. Generation destroys composited detail, so anything painted, comped, or type-set has to happen after the final generative pass on that frame. In practice that means generating a surface deliberately blank and comping the artwork onto it afterwards, rather than asking a model to put the right image in the right place and hoping. Get the order wrong and you pay for the same frame twice.

The DistroKid stack in order: ComfyUI generates, Photoshop cleans up by hand, Seedance generates motion, then After Effects and Premiere Pro finish by hand. Nothing regenerates after the Seedance step. COMFYUI Flux + trained LoRAs 01 GENERATIVE PHOTOSHOP cleanup, correction 02 BY HAND SEEDANCE first/last frame, i2v 03 GENERATIVE AFTER EFFECTS composite, post 04 BY HAND PREMIERE PRO sound design, edit 05 BY HAND LAST GENERATIVE PASS ABOVE THIS LINE RAW MATERIAL THE SHOT
The stack, in order