The
Consistency Problem

A generative model is built to reinvent. Ask it for a face and it will give you a plausible face, sampled fresh each time from everything it has learned a face can be. That is exactly what you want when you are exploring, and exactly what breaks a commercial. A brand does not need a plausible face. It needs the same face, the same product, the same logo and the same colour in shot 1 and in shot 40. The gap between those two demands is the consistency problem, and most of the real work in AI production lives inside it.

01 - The DriftWhy models forget what they just made

The reason drift happens is structural, not a bug to be patched. A diffusion model does not store your character. It stores a distribution, a vast statistical sense of how pixels tend to arrange into people, rooms, objects. Each generation is a fresh draw from that distribution, nudged by your prompt and your seed. Nothing in the process carries an identity forward. So when you ask for the same woman in a new shot, the model is not recalling her. It is inventing a new woman who satisfies the same description. The description is coarse. "Thirty, red coat, dark hair" fits a million faces, and the model will happily use a different one each time. What you experience as forgetting is really the absence of memory in the first place.

02 · Two Kinds of SameLook-consistency versus identity-consistency

It helps to separate two things clients tend to merge. Look-consistency is the easier one: the same grade, grain, lens character, lighting logic across a sequence. That is largely a matter of prompt discipline, reference framing and colour work in the grade. Identity-consistency is the hard one: this exact person, this exact bottle, this exact wordmark, holding across angle, motion and lighting change. Models are reasonably good at the first because a "look" is itself a distribution, a mood, a palette. They are poor at the second because identity is a single point, and a model that samples a region will keep missing a point. Most consistency failures are someone hoping look-consistency will quietly deliver identity-consistency. It will not.

03 - The ToolkitHow far you can push generation

There are ways to push a model toward a stable result: references, conditioning, fine-tunes, seed discipline and control inputs. They help, but they do not remove the underlying distinction between look-consistency and identity-consistency. The practical node-based setup belongs in Why Node-Based Control Beats Prompting.

04 · When Generation Can't Hold ItRebuild in 3D or composite the real thing

There is a threshold where the honest move is to stop asking the model to remember and give it something it cannot forget. On the Gigaclear national spot we needed a crowd of a hundred characters that stayed the same people across every angle of the edit. No amount of seed discipline survives a hundred identities through a camera move. So the crowd was built once in Cinema 4D and then controlled, the same geometry, the same materials, viewed from wherever the shot required. The identity was authored, not sampled, which is why it did not drift. Live actors were shot and composited in over that base. That is the pattern: when generation cannot hold identity, you rebuild the asset in 3D where identity is a property of the model, or you composite a real element that was never in question. The generator becomes one layer in a shot, not the whole shot.

05 · Products, Logos and TextThe things AI will not respect

This is where clients are most exposed and least warned. A product must be pixel-correct, the client's bottle, their exact label, their trademarked mark. A model has no concept of a trademark. It will render something in the neighbourhood of the logo, subtly wrong in proportion and kerning, and it will do so with total confidence. Typography fails the same way and for the same reason: the model is drawing the shape of text, not spelling it, so a tagline arrives as convincing gibberish. The rule I work to is simple. Anything that must be legally or factually exact, logo, packaging, product geometry, copy. Is not generated. It is supplied as a real asset and composited, or reconstructed in 3D and tracked in. You let the model make the world and you place the trademark into it by hand.

06 · A Pipeline ProblemThe useful part is the system around the model

The mistake is to treat consistency as something you can prompt your way into. You cannot, because the failure is baked into how the model works. It samples, it does not store. Consistency is won in the system around the model: the reference libraries, the trained LoRAs, the ControlNet rigs, the 3D assets built once and reused, the compositing stage where real logos and real faces are laid back in, the discipline of changing one variable at a time. None of that is the model. All of it is the pipeline. A model on its own will give you forty beautiful shots of forty different people. A pipeline gives you the same person, forty times. For a brand, that difference is the entire job, and it is why the useful part of this work has never been the model. It is the system built around it.