Brief
Brand study from the live platform. Tone pillars and a hard avoid-list set before any generation.
Two complete :15 spec campaigns, fourteen final frames and two films, produced solo through an AI-assisted pipeline under human art direction. Stills and motion generated through APIs, beat-cut in code with Remotion.
The test: study a real brand’s tone, product line, and visual language, then deliver two distinct, finished social campaigns using current image and video models, with every creative decision made by a human art director.
The hard part of AI production is not generating images. It is consistency: the same camera, the same product colorways, the same character, frame after frame, until nine separate generations read as one campaign. That is a discipline problem, not a prompting problem.
The bar for every frame
“It has to pass as real, or it doesn’t ship.”
Project A
One fixed top-down camera. Feet in Crocs, dead center, every frame. Everything else changes. Nine ordinary moments from one life, beat-cut into a :15 loop. The locked composition is what makes nine generations read as one campaign.
Project B
One character, one fixed frame, five looks. The Crocs change with every outfit. Hard cuts land mid-gesture, and the styling coordination is the content. Character consistency held across every still and clip from a single locked face reference.
Brand study from the live platform. Tone pillars and a hard avoid-list set before any generation.
A product reference library from catalog photography, colorways checked at the pixel level.
One hero frame iterated to approval, then locked as the composition reference for every frame after it.
Generated against the locks and checked at zoom per select. Anything that does not pass as real is cut.
Micro-motion keeps subjects still while environments move. Identity refs ride along on every clip.
Cut grids built in code from each track’s measured BPM. One global grade across every select.
The approved frame-one image was attached to every later generation as a composition reference. Camera angle, product scale, and stance had to match it exactly. When a graphic ground kept pulling the pose, the lock won.
Project B’s character came from a single face reference shot on mid-gray seamless, because white backdrops bleed into final lighting and push skin toward plastic. Any marker that drifted between looks was cut from canon rather than patched per frame.
Editing models stay close to their base image: they remove elements and clean surfaces but resist repositioning. Re-render for layout changes, surgical edit for local fixes. Every select checked at full resolution, because feed-size review misses seating and mount errors.
AI flash-swaps and morph transitions both read as AI: one felt wrong on the beat, the other hybridized mid-frame. Every cut in both films is a straight cut in the Remotion edit, with Project B’s cuts landing mid-gesture.
The first motion pass showed face drift and AI skin, because image-to-video only sees the plate. The fix was reference-to-video at 1080p with the character canon attached to every generation. Upscalers sharpen what is there; they do not fix it.
Anyone can generate a good-looking AI image. The difference between a demo and a campaign is whether frame seven matches frame one: same camera, same colorway, same charms in the same holes. That takes reference discipline, a requirement that every frame pass as real, and an art director willing to throw out good frames that do not match.
The pipeline produced two campaigns for one brand, with a full generation log behind every select. The process generalizes to any brand.
Brief to campaign
Another AI production workflow