A 3D render knows the geometry. Model a knurled cap once and it returns the same knurl from every angle, at every focal length, for as long as the file survives. A generative model does not know it. It infers — convincingly on the first frame, less so on the ninth.
That one difference decides most of what goes to which tool. CGI for the object. AI for the world. Then the part that is neither: the frame where the two meet.
We don't generate. We direct. In practice that means deciding, shot by shot, which half of the picture is measured and which half is imagined.
What a render is better at
Geometry a client can check. A fragrance house knows its own bottle to the millimetre — neck ratio, shoulder radius, the height the label sits at. A render holds those numbers. A generation approximates them, and the approximation is usually flattering, which is worse than wrong.
Repetition is the second reason. RIIFFS runs two lines and more than forty products, NUSUK on the heritage side and RIIEFF on the designer side. A catalogue that size needs the same camera move on every product, and 3D rendering gives you that for nothing. Ask a model for the same lockup forty times and you get forty cousins.
The third reason is invisible and matters most. A render arrives in passes — depth, matte, specular, ambient occlusion, an alpha with edges that were computed rather than guessed. Those passes are the working currency of compositing. A generated frame arrives as flattened RGB and nothing else.
What a generation is better at
Everything the object is standing in. Weather, dust, a corridor with three centuries on its walls. Building that in 3D is weeks of set dressing. Describing it is an afternoon.
It is also better at mess. The failure mode of a CGI environment is that it looks new. On COLOSSEUM the rule went into the direction before a frame existed: no surface is CGI-clean. Hairline cracks in the plaster, dust on frame edges, worn marble thresholds. A generative model produces that decay unasked, because decay is most of what it has seen.
Ramón Béjar runs on one written rule — No Faces, Only Storytelling — and not a single human face appears in the library. Architecture sets the frame and light is built for material, so every scene has to arrive on surface and colour instead. A generated environment is the quickest way to build architecture that does that job. A render is the quickest way to put the bottle inside it without softening an edge.
And it is better at the scale of a world. RIIFFS needs desert and heritage behind one line, dark-luxe behind the other. The engine builds the world behind the bottle. Direction decides which of the two it is.
The seam
AI compositing is mostly translation. The generated plate has no camera, no depth and no alpha, so the first job is to invent all three credibly. Estimate depth, commit to a focal length, set a horizon, then drop the render into a lens that never existed.
Then the matches, in order, because the order is not negotiable. Light direction first — get that wrong and nothing downstream rescues it. Colour temperature second. Contact shadow third, and this is where most composites die. An object that is not touching anything reads as a sticker, however good the render is.
Grain and lens last. A render is too clean and a generation carries its own texture, so both get taken to one film response: the same emulation, the same aberration at the edges, the same falloff. If you can tell which half was measured, the seam has won.
Reflections are the honest hard problem. Polished glass and brushed metal record their surroundings, so an object lit under one scheme keeps reflecting that scheme after you replace the world behind it. Sometimes the answer is a matched HDRI on the render. Sometimes it is choosing the world before you build the light.
Motion multiplies all of it. A still composite has to survive one frame. A moving one has to survive parallax, and a generated plate that redraws its own architecture between frames will not hold a tracked object. So the camera move is decided first and the plate is built to serve it, never the other way round.
When not to composite at all
L'Orientale Fragrances gets two backgrounds and one house. Dark and cinematic for the campaign — gold and amber on matte glass, shadow doing most of the work. Then the same bottle on white with nothing behind it for the product grid.
The white version has no seam to hide and no world to invent. Its discipline is a different one: the art direction has to survive the jump. That is a judgement about which frame needs a world, not a preference about which tool is better.
The seam is where the craft is because it is the only part neither tool does for you.