Photorealism Has Three Orthogonal Axes and Most Workflows Only Fix Two

24Latents · reddit · 2026-08-26

Based on months of hands-on work with Krea 2 + character LoRA, the author argues photorealism in AI images has three orthogonal axes: optics (fight the too-perfect camera with grain, flare, imperfect exposure, tilt), scene (fight the showroom set with clutter, wear, off-axis furniture), and subject pose — the axis almost nobody touches, and the reason technically correct renders still read as fake.

Key claim: style tokens like candid documentary photo shift rendering, not pose. Without a motivated action, the model falls back to its most photographed pose: frontal, centered, symmetric, eyes to lens. A character LoRA makes this worse by stacking a portrait prior on top.

Five levers, strongest first:

A gesture toward the world: write concrete physical markers (caught mid-stride, one foot planted ahead), never bare verbs — and never lists of states ("weight on one hip, shoulders dropped"), which give the model nothing to build a pose from.
Self-directed gestures: one hand only, fatigue register only (kneading neck, arm pressed to ribs). Banned outright: both hands in hair (pin-up prior), arms crossed over chest (studio self-embrace), any back arch.
Geometrically compatible anchored gaze: a turned head needs an eye target inside the head's facing cone; otherwise the model silently reconciles, drops the gaze anchor, and eyes land on the default target — the lens.
Break the body: collapsed weight on one hip, dropped shoulders, asymmetric stance; never feet wide apart on a static figure (reads as monumental sculpture).
Candid composition: explicit off-center placement, slight lateral crop, subject partly eaten by shadow.

Original post →

Photorealism Has Three Orthogonal Axes and Most Workflows Only Fix Two

More from Multimodal