ChatGPT
Perplexity
Claude
Gemini
Nobody trusts a perfect picture anymore. For a while now, the pitch behind every new image and video model has been the same: sharper, cleaner, more photoreal, less broken. Resolution keeps climbing, faces hold together across more scenes than they used to, and lighting looks correct in a way it didn't two years ago. And somewhere in the middle of all that progress, something strange happened: technical perfection stopped being the flex.
Los fotógrafos se toparon con esta misma tensión antes de que lo hicieran los modelos generativos, y la dirección que tomaron es el anticipo más claro de hacia dónde se dirigían la generación de imágenes y vídeos. Alex Cooke, de Fstoppers, lleva tracking a shift among working photographers away from flawless retouching and toward images that keep their rough edges, things like a missed focus point, an unretouched tear, a hand slightly blurred mid-motion. His argument is that AI didn't just get good at photography, it got good at the specific kind of photography that used to signal skill. Once software can smooth skin, fix a horizon, and match color across a hundred images in the time it takes to make coffee, flawless execution stops signaling that a professional made something and starts signaling that it might have been generated instead. Cooke's point isn't that photographers should get sloppy; it's that the thing a viewer used to read as competence now reads as suspicious, mientras que lo que antes se leía como un error ahora se lee como proof alguien estuvo realmente allí.
Algo parecido está ocurriendo dentro de la propia generación de imágenes con IA, un nivel por encima: a medida que los modelos mejoran a la hora de representar la versión objetivamente correcta de una cara o una escena, también mejoran a la hora de parecer que no muestran nada en concreto. Es esa cualidad concreta, vidriosa y demasiado simétrica que la gente ha empezado a reconocer al instante como basura de IA. Dominica Baird, directora del departamento del programa de negocio de la belleza y la fragancia de la SCAD, señalado that when photography was invented, painters spent a while trying to out-realism the camera before they gave up on that fight. They moved instead toward abstraction, symbolism, and expression: toward everything a camera couldn't do. That's roughly the same shift happening now, one rung up the ladder, as AI absorbs "technically correct" and craft migrates toward the specific, chosen, slightly imperfect thing a model wouldn't generate on its own unless you told it to.
What's actually happening underneath the aesthetics is less about how something looks and more about what it proves. An image with a slightly off crop, a grain, a beat that runs half a second too long, reads as the result of somebody making decisions: a pass, then another pass, a choice to leave something in rather than smooth it away. Slop doesn't feel like slop because it's ugly. Plenty of slop is technically flawless. Parece un churro porque se lee como el resultado de one click, una única generación, por mucho cuidado que se haya puesto en el prompt que hay detrás. The imperfection is what tells a viewer that more than one decision happened between the idea and the thing in front of them, and that gap, between a single generation and a shaped one, is the actual difference between work that earns a few seconds of someone's attention and work that doesn't.
El mismo mecanismo está empezando a aparecer también en el vídeo, solo que un paso por detrás de donde ya se asentaron las imágenes. La fotografía y la generación de imágenes lo resolvieron primero, y el vídeo aún busca su sitio, pero avanza lo bastante rápido como para que el mismo lenguaje ya esté apareciendo este año en la cobertura del diseño de movimiento. Envato's own 2026 trend report names "authenticity through imperfection" as one of the defining shifts in video right now: handheld camera sway instead of gimbal-smooth motion, raw cuts left in rather than trimmed away, natural pauses and background noise kept in the mix, wearable POV footage shot with "no rigs, no setups." Envato's read on why is blunt: "imperfection has become its own aesthetic. It signals trust, relatability, and authenticity."
Incluso las empresas que desarrollan herramientas de vídeo con IA han llegado a la misma solución desde dentro. Hridaye, el director creativo de invideo, describió su proceso como añadir un toque de desenfoque y una pasada de grano hasta que un clip generado se parezca más a una película de acción real, y el razonamiento coincide casi exactamente con el argumento de Cooke sobre la fotografía. Como lo explica el propio invideo, los fotogramas de IA "no tienen ruido de sensor, ni estructura química, ni imperfección orgánica", y esa ausencia es "lo que se percibe como 'falso' antes de que puedas articular por qué".
None of this works as a checkbox, though, and it's worth being honest about where the idea breaks. Communications strategist Shaunta Garth has defendió lo contrario about brands manufacturing imperfection on purpose: the moment roughness becomes a formula, applied because a trend report said to, it stops being evidence of a decision and turns into just another preset, exactly as hollow as the smoothness it replaced. A grain overlay dropped onto an otherwise untouched first generation is still a one-step output, just wearing a different filter, since the imperfection itself was never el verdadero objetivo. La iteración detrás sí lo era.
Which is the actual craft lesson, for images and video both: the goal isn't imperfection for its own sake, it's a body of decisions a model wouldn't have made on its own. Describe the conditions of capture instead of the result, not "a beautiful portrait" but the specific lens, the specific light, the specific flaw you want in the frame, and do the same with motion: not "cinematic camera movement" but the sway of a hand actually holding something, the pause before someone speaks, the cut that runs a beat long because that's how it happened. Push past the first clean output. The second or third pass, where you start pulling the piece away from its most polished default, is usually where the real decisions live, and that's true whether you're working in GPT Image 2 and Nano Banana Pro or Kling 3.0 and Seedance 2.0.
None of this is a rule about what AI work should look like. Plenty of it should be clean, bright, and precise, and that's a legitimate choice too, as long as it's made the same way: on purpose, after a few passes, not on the first try. The point was never that clean is wrong. It's that clean is no longer proof of anything, because a model can hand you clean without you doing a thing. The choices that still require you, esos que una sola generación no produce a menos que vuelvas a entrar y los fuerces, es donde el trabajo empieza a parecer de alguien en concreto, y donde el público decide que merece la pena verlo.