Testing text-to-image prompts across multiple themes revealed clear strengths alongside definite technical boundaries. We subjected Dream Lab to three specific creative briefs: technical product renders, multi-subject human portraits, and complex typographic badges.
Product mockups fared best. Prompts requesting "matte ceramic cosmetic bottles on fluted marble pedestals with sharp side shadows" generated clean specular highlights, physically plausible liquid refractions, and straight geometry. The engine handles hard surfaces without the wobbly warping common in early diffusion tools.
Prompt Test 1: "Minimalist skincare serum bottle on travertine slab, architectural shadows, photorealistic, 8k"
Result: Zero edge fringing, pristine product geometry, subtle surface grain intact.
Prompt Test 2: "Crowded outdoor subway platform at dusk, commuters checking phones, rainy reflections"
Result: Foreground figures sharp and natural; background anatomy shows minor blurring past 30 meters depth.
Human anatomy showed marked improvements over previous iterations. Five-fingered hands emerged consistently across 18 of 20 test generations. Subtle dermal details, like skin pores, stray hairs, and natural eye reflections, rendered cleanly without that slippery, wax-like sheen typical of unchecked sub-surface scattering algorithms. Distant background crowds still exhibit occasional loss of facial coherence, but foreground subjects hold up under close inspection.
Native typographic accuracy inside Dream Lab still encounters hiccups. While short words like "FRESH" or "SALE" render reliably within geometric badges, strings exceeding four words frequently suffer from duplicate letters or dropped consonants. For commercial design workflow demands, typing live vector type over an AI-generated blank plate remains the professional approach.
