Nano Banana 2: Google's latest AI image generation model

Original Article ↗ Hacker News Discussion ↗

Model Positioning & Naming

  • Confusion around branding and versions: “Nano Banana 2” corresponds to gemini-3.1-flash-image-preview, distinct from Gemini 3.1 Flash text models and from “Pro” image models.
  • Some want Google to drop the banana name; others argue the quirky branding has strong recognition.
  • Model card benchmarks suggest NB2 is close to or slightly better than Pro on overall preference/visual quality, but not a clear leap.

Capabilities, Quality & Performance

  • Users report strong photorealism and architecture/interior outputs; NB2 can handle some prompts that previously stumped SOTA models.
  • Weak spots: layout/structured prompts (e.g., 5×2 grids), editing localization (tends to change too much of the image), transparent PNGs, and handling truly novel “A with feature X from Y” combos.
  • New behaviors include detecting conflicting instructions in prompts and configurable “thinking” levels.
  • Generation is often slow (2–3 minutes), with occasional “resource exhausted” errors; some say quality gains over NB Pro are incremental, not transformative.

Pricing & Access

  • NB2 is significantly more expensive per 1024×1024 image than original NB, but cheaper than NB Pro, with new tiered pricing up to 4K.
  • Some speculate older, cheaper models may eventually be deprecated; others note Google defaulting UI to “Fast” (cheaper) hints at GPU constraints.

Practical Use Cases

  • Heavy use for rapid ideation: drafts, mockups, storyboards, diagrams, and stock-art replacement.
  • Several detailed workflows for house building, landscaping, and café design: users feed SketchUp/floorplans/photos into NB to iterate, then hand renders to draftsmen, cabinet makers, or contractors.
  • Tools are starting to displace mid-tier interior designers and illustrators in budget-constrained contexts.

Impact on Creative Work & Jobs

  • Ongoing argument whether this is just “we don’t want to pay artists” or a legitimate productivity boost akin to photocopiers and email.
  • Many foresee erosion of low-end commercial art/stock work; disagreement over how much high-end illustration, branding, and art direction will survive.
  • Some see new roles emerging (prompt specialists, taste-curators), others doubt they’ll offset losses.

Misinformation, Porn & Trust

  • Strong concern about deepfakes, fake OnlyFans models, scams, and political disinformation; many believe most internet users have already mistaken AI images for real.
  • Debate whether ubiquity of fakes will improve media literacy or simply increase polarization and cynicism.
  • Porn-specific discussion: open models already strong, censorship on commercial models is porous, and people expect highly personalized parasocial porn systems with serious exploitation potential.

Art, Originality & Cultural Value

  • Long philosophical thread: does AI art merely remix training data, or can it be genuinely creative?
  • Some argue art requires lived experience and embodiment; others claim humans also “just remix,” and future models could approximate “taste” via RL from expert curators.
  • Analogies drawn to photography’s arrival and Walter Benjamin’s “aura”; many predict physical, analog, and live art (sculpture, film, concerts) will gain relative prestige as digital images become cheap and ubiquitous.

Changing Relationship to Images

  • Many feel emotionally numbed by the flood of perfect images; compare it to smartphone photography diluting the specialness of rare film photos.
  • Counterpoint: curation, context, and personal connection still create emotional weight (family photos, physical prints, film, Polaroids).
  • Prediction that “AI slop” will dominate mass content, while authenticity, imperfection, and visible human effort become premium signals.