Natalia Farsang-Bajczi — AI Technical Artist at OpusClip
I build the style systems behind AI video.
A style isn't a look — it's a contract. It has to hold across thousands of inputs nobody planned for, inside a pipeline I only partly control. That takes an eye and a working model of the whole system.
What the job actually is
Taste is the input. A system is the output.
- 01
Define the look
Art direction that has to survive contact with footage I have never seen, at sizes I did not choose, with nobody in the loop to fix it.
- 02
Find where it breaks
A trained eye tells you something is wrong. It does not tell you where. Locating the stage actually responsible is a different skill, and it is most of the job.
- 03
Make it a system
Turn the look into a spec with stated guarantees and named limits — something another person can pick up and use correctly without me in the room.
Source → generated
What goes in, and what comes out.
One video-to-video pass through Seedance 2.5. On the left, the reference that went in: a phone clip, available light, nothing staged. On the right, what came out — same subject, same product, an entirely new set of shots and a grade that holds across all of them.
The interesting part isn't that it looks better. It's that the look has to survive every shot, not just the one you cherry-pick — which is the whole reason a style needs to be a spec rather than a prompt.
Input
The original video
Shot on a phone. Available light, one take, nothing staged.
Output
Generated with Seedance 2.5
New angles, new lighting, one grade holding across every shot.
Selected work
All case studiesA style is a contract
Turning one art-directed look into a spec that survives thousands of inputs nobody planned for.
- Style systems
- Consistency
- Evaluation
Keyframes as a control surface
The one place in a generative video pipeline where an art director still has direct authority.
- Keyframes
- Continuity
- Composition
Range
Range is the evidence.
Generated stills, each written up as a spec rather than shown as a picture: what the look is for, what it guarantees, and the conditions under which it falls apart. The last part is what makes it a contract instead of a filter.
The Artifact Atlas
A field guide to how AI video breaks.
Everyone publishes the shots that worked. This is the other half: named failure modes, what causes each one, and what actually fixes it — with clips showing the artifact rather than describing it.
Read the atlas- 01Temporal flickerFrame-to-frame instability in brightness, texture or fine detail that is invisible in a still and obvious the moment the clip moves.
- 02Identity driftA subject's face, build or defining features change gradually over a clip, so the first second and the last second are recognisably different people.
- 03Style bleedA style applies unevenly — leaking into regions it should not touch, or resolving differently in adjacent shots that were supposed to match.
- 04Keyframe seamsA visible discontinuity exactly where two generated segments meet — the one moment in a clip when a viewer is already paying close attention.
Notes
All notesA style is not a prompt
A prompt describes an intention. A style has to make a guarantee — and the difference shows up the moment someone else uses it.
Where style actually lives in a pipeline
The stage where you can most easily change a look is rarely the stage where the look is actually decided. Knowing the difference is most of the job.
Contact
If your video pipeline looks right in stills and wrong in motion, I can help.
Earlier, in advertising
- Disney
- National Geographic
- Discord
- Capcom
- Nivea
- Pampers
- Starbucks
- Amazon
- Oral-B
- Beats
- Asahi