Maya.
One real photograph, rebuilt into a complete reference canon. Every asset is scored against that source before it ships — the score sits on each tile, and where a score cannot be trusted the tile says so.
Maya photographed in her before state. Same pipeline as the after shoot — two face references from the before canon, one real face-free world reference, the capture layer — with the styling register re-pointed: ordinary specificity rather than elevated, and a fitted wardrobe so her figure is plainly visible.
The reference set these are built from lives in Canon.
Each pair is the same circumstance in both states — the same errand, the same mirror, the same car, the same night out. That pairing is the point: a transformation reads when the situation is held constant and the only variable is her.
Both panels are matched to a common height and given identical treatment. No arrows, no burned-in labels, no divider graphics — the captions sit outside the image area so the photographs stay photographs.
Both states trace to the same origin. The after canon and the before canon are each built star-topology from profile.jpg, so neither is derived from the other and neither inherits the other's drift.
profile.jpg
The single confirmed photograph of Maya, and the sole origin of every reference below — after and before alike. Nothing else feeds identity.
The previous Maya Biometric/ sheets were rejected outright: they measured 0.21–0.36 against this image, the same band that condemned the legacy set as a different person. They were studio glamour renders, not photographs.
Star topology. Every reference is exactly one edit from this file, never edited from another render. The before canon is built from this photograph directly rather than from the after canon, so the two states are siblings and neither inherits the other's drift.
The collage is assembled locally from the exact nine master files. It is never re-rendered as a grid — asking a model for a 3×3 would reinterpret all nine into slightly different people.
2K, never 4K. At 4K the output is upscaled somewhere in the backend and skin turns mushy and faintly cartoonish. The forehead lines in the snarl and surprise masters, the nasolabial folds, the uneven complexion — that texture is what 4K averages away.
Both profiles were generated independently, not mirrored from one another, so they are genuine opposite views of the same inferred head.
Why the profiles are not gated. ArcFace similarity correlates with face width in frame — it substantially measures how large and how frontal a face is, not whether it is her. A true 90° profile presents almost no frontal information, so gating it would reject correct geometry and pull the set back toward flat frontal frames. Frontal gates at 0.70, three-quarter at 0.60, profiles are reported and verified by eye. The eye decides.
Known gap: a5_profile_right sits closer to 80° than a true 90°. Declared rather than quietly accepted.
Every gate above is a floor: more similarity to the source is better. That reverses the moment the brief is a deliberate change to the person. A before tile scoring 0.85 against the source does not mean success — it means the weight never applied and the after canon came back under a different filename.
So the before canon gates on a window, and the two failures point in opposite directions: below the floor the change overshot and it is a different woman; above the ceiling the change never happened. A one-sided gate is structurally blind to the second failure, and the second failure is the one that looks like success.
It caught two immediately. chin_down returned 0.842 / 0.807 / 0.861 and soft_smile 0.807 / 0.800 / 0.824 — six consecutive takes a floor-only gate would have ranked as the best in the set. Both failed for the same anatomical reason: a closed-lip smile lifts and tightens the cheek, and chin-down compresses the under-chin into the neck. Those are exactly the regions where added weight shows, so each pose hides its own evidence. They are declared as gaps rather than forced.
Weight is soft tissue and cannot move bone. Cheek fullness, jaw softness, under-chin, neck and lid weight are free to change. Interocular distance, the nose bridge line and length, brow arch, hairline, ear position, lip outline, freckles and her natural asymmetry are load-bearing identity and survive untouched. That split is what keeps her recognisable through a large change.
Escalating the ask failed, and then backfired. Requesting fifty pounds with a longer, more emphatic magnitude block produced less change than requesting forty with a short one — 26 of 28 attempts landed above the ceiling, scoring higher than the pass before. The rhetoric buried the number. That is this system's own first law turned on its author: the instruction got louder and more diluted at the same time.
So the mechanism changed instead of the wording. The model resists one large change away from a reference face, and no amount of asking fixes that because asking means writing more. Two moderate steps travel further than one large one, since at each step the change is small relative to its own reference. The before canon therefore takes one bounded hop: profile.jpg → a first heavier tile → these. One hop, never a chain, and every tile is still scored against the original photograph so the generation loss is visible rather than assumed.
The body needed its own hop. These canon tiles are head-and-shoulders — they carry face weight only. The body in every shoot frame came from a prompt clause, so a perfect face canon could never have fixed it. A second hop is applied to the finished photographs, where the reference is a full body in a real scene and the scene, clothes, pose, light and framing are all held constant. That is where the waist, arms and thighs finally moved.
A caution carried forward. ArcFace similarity does not track facial weight — it tracks pose, framing and light. Tiles that were visibly fuller scored higher and were wrongly rejected before that was caught. The window is a coarse sanity check on this page, nothing more. The eye decides.