The tells that still betray a fake in 2026, ranked by how often I catch them

by @avaalmeida_fx in c/photorealism · · Critique Me

Been doing blind-test threads for a year. The skin/hands era is mostly over. Here's what actually gives it away now, most→least common:

Micro-text and logos. Any writing in the scene — a street sign, a book spine, a label — is mush or dream-letters. Instant kill. #1 by a mile.
Specular/reflection geometry. Reflections that don't match the scene, catchlights that don't match the environment, a window reflecting a room that isn't there.
Physical implausibility in contact. Where things TOUCH — fingers on a cup, feet on ground, a bag strap on a shoulder — the pressure/deformation is wrong. Nothing dents, nothing bears weight.
Depth-of-field that's painted, not optical. The bokeh is a blur filter, not a lens. Highlights in the background don't bloom into real bokeh balls, and the focus falloff isn't continuous with depth.
Too-perfect symmetry/repetition. Identical bricks, identical leaves, both eyes identical. Reality is noisy.

Fix text by inpainting it out or hand-replacing it. Fix reflections with a 3D proxy or by choosing matte subjects. Fix contact by composing around it. There's no prompt for 'obey physics'.

1,181 points · 9 comments

Comments

  • @oscarcostacreative · · 97 points

    Contact/pressure is my favorite one because it's subtle and nobody's trained to look for it until you point it out. A hand 'holding' a mug where the fingers don't actually wrap or press. The mug would fall. Dead giveaway once you see it.

  • @hanamillerlab5022 · · 70 points

    The weight thing generalizes — nothing in AI images has WEIGHT. Clothes don't pull, cushions don't compress, a heavy bag doesn't dig into a shoulder. Everything floats in a gravity-free pretend.

  • @quinnjackson_art · · 68 points

    Would add #6: light that comes from nowhere. A beautifully lit face with no visible or implied source, or shadows going three directions. Our brains flag impossible lighting even when we can't articulate why.

  • @kaibrowncreative · · 98 points

    Text is STILL the #1 tell and it's wild how far everything else has come while legible signage remains a disaster. Diffusion models fundamentally don't do symbolic sequences, they do texture-of-text.

  • @rileypark_studio · · 99 points

    'texture-of-text' is exactly it. The model knows what text LOOKS like (a row of marks) not what it IS (a sequence of symbols). It's painting the vibe of letters.

  • @diegomoore_vfx · · 52 points

    optical vs painted bokeh is the one I use to catch 'photographers' faking it. real bokeh has cat-eye swirl at the edges, onion rings in the balls, longitudinal CA. the blur filter has none of that, it's gaussian mush.