As IA conseguem gerar imagens fotorrealistas de rostos humanos não identificáveis como sintéticos ?

A IA generativa transformou a criação de media visual. Ferramentas como Stable Diffusion e DALL-E produzem agora imagens indistinguíveis da fotografia para observadores humanos. Elas tratam a iluminação, textura e nuances emocionais com um realismo surpreendente. Isto difumina a linha entre o real e o sintético, desencadeando debates sobre autenticidade e desinformação.

Background

Generative AI has transformed visual media creation. Tools like Stable Diffusion and DALL-E now produce images indistinguishable from photography to human observers. They handle lighting, texture and emotional nuance with uncanny realism. This blurs the line between real and synthetic, sparking debates on authenticity and misinformation.

Current systems like DALL-E 3, Midjourney v6, and Stable Diffusion XL can produce highly photorealistic face images indistinguishable from real photos to most human viewers, especially when guided by prompt refinements such as “ultra-realistic, 8k, subtle skin texture, natural lighting.” These models achieve this fidelity by training on large datasets of licensed portrait photographs while incorporating adversarial filtering and diffusion-based refinement to suppress obvious synthetic artifacts. However, even state-of-the-art generators still exhibit subtle inconsistencies—such as improbable reflections, skin-pores that misalign with lighting, or teeth arrangements that violate anatomical norms—that can be detected under close inspection or by forensic analysis tools.

— Enriched May 12, 2026 · Source: Adobe Research