Services

Generative and AI Visual Direction

AI and generative pipelines used as art direction, not as a prompt lottery.

The difference between an AI visual that works and one that looks like everyone else's is almost never the model. It is dataset work, curation volume, and someone making art direction decisions across thousands of outputs.

Ötme Bülbül Ötme landed in early 2023, before usable commercial AI video tools existed. Stable Diffusion 1.2 produced muddy output, so custom dataset training — including on the performer's face — became necessary to keep portraits anchored to the music instead of drifting into generic synthetic faces. Roughly six months of synthesis and curation produced more than five thousand images and a thousand video generations, shaped into a single visual line.

That is what this service actually is: building and curating at scale so the result reads as one continuous world rather than a collage of experiments.

What you get

  • Visual direction and reference development for the generative language
  • Custom dataset preparation and model training where the look requires it
  • Image and video synthesis at production volume
  • Interpolation, compositing, and integration with recorded or stock footage
  • Curation — the selection pass that decides what the piece actually is
  • Delivery in the format your edit or installation needs

How it runs

  1. 01

    Visual language definition

    What the piece should look like and, more importantly, what it should not. Generative work without this becomes noise very quickly.

  2. 02

    Pipeline and dataset setup

    Model selection, dataset preparation, and training where a specific subject or look has to stay consistent across thousands of frames.

  3. 03

    Synthesis at volume

    Generating far more material than the final piece will use, because selection is where the quality actually comes from.

  4. 04

    Curation and assembly

    The selection and sequencing pass, plus compositing against recorded or stock material where the piece needs an anchor in something real.

  5. 05

    Delivery

    Handover in the form your editor, installation, or broadcast pipeline requires.

Good fit for

  • Music videos and artist collaborations with a distinct visual world
  • Installation content where generative material needs internal consistency
  • Brand films looking for a look that is not achievable with stock or standard 3D
  • Projects with enough schedule for curation — this work rewards volume

Not a good fit for

  • Fast-turnaround content where AI is a cost-cutting shortcut rather than a direction
  • Work requiring exact likeness rights or documentary accuracy
  • Projects where the brief is only 'make it look AI'

Ways to work together

Where this has been done

Public projects that show this work in practice, with credits and venues.

Mindflow

Nine-channel audiovisual installation combining AI, motion capture, and live improvisation into a synchronized spiral of video and sound.

View case study →

Common questions

Is this just prompting?

No, and projects that treat it that way tend to look like it. The work is dataset preparation, training where consistency demands it, generating far more material than gets used, and then curating hard. On Ötme Bülbül Ötme that meant over five thousand images and a thousand video generations for one music video.

How long does AI visual work take?

Longer than people expect, because volume is the mechanism. Six months went into the synthesis and curation on Ötme Bülbül Ötme. Modern tools have compressed that considerably, but a piece that needs a specific, consistent world still needs weeks rather than days.

Have something like this coming up?

Send the rough scope, the venue, and the dates. If it is not a good fit I will say so, and where I can I will point you to someone better placed.

Start a conversation