Nano Banana Pro

Original Article ↗ Hacker News Discussion ↗

Model capabilities and limitations

  • Many commenters find Nano Banana Pro a big leap over prior Google image models, especially for:
    • Legible, accurate text in images and infographics.
    • Localized image editing (e.g. removing specific objects, changing small details) with minimal collateral changes.
    • Compositional prompts (multiple constraints in one scene) and UI/infographic/layout generation.
  • Benchmarks shared in the thread show clear gains over the original Nano Banana, especially on tricky editing tasks (e.g. block‑moving “SHRDLU” tests, selective object removal).
  • Still-visible weaknesses:
    • Structured content like piano keyboards, maps, sprite sheets, road diagrams, and some architectural/physical layouts.
    • Style transfer for specific artists (e.g. Studio Ghibli) and consistent faces in edits.
    • Transparency/alpha and fine animation sequences remain rough.
    • Output sometimes varies wildly between runs for the same prompt.

SynthID watermarking and authenticity

  • SynthID is described as an invisible watermark embedded in images to signal “made by Google AI”, now checkable via the Gemini app.
  • Several people note:
    • EU rules push big providers toward watermarking, but no standard method or robustness is mandated.
    • Tools and papers already claim high success rates at removing SynthID‑style watermarks; some share open-source removal projects.
    • Watermarking only proves that some images are AI; absence of a mark doesn’t prove authenticity, especially with open or “grey market” models.
  • Broader concerns:
    • Risk of conditioning people to trust “no watermark” as real, making sophisticated fakes more persuasive.
    • Potential for surveillance and legal overreach if platforms must check every upload against proprietary detectors, possibly uploading all user media to major AI vendors.
    • Counter‑proposals include camera-side signing (C2PA) and visible watermarks for everyday users.

Rollout, pricing, and UX

  • Rollout is described as vague and fragmented: partial availability in Gemini app, AI Studio, Vertex AI, and third-party platforms.
  • Many report confusing “permission denied” errors, unclear model names, and especially painful billing/onboarding through Google Cloud.
  • Some circumvent this via aggregators (e.g. Fal, OpenRouter), which present simpler APIs and clearer pricing.

Use cases and impact on creative work

  • Reported uses: marketing images, blog illustrations, memes, research diagrams, product mockups, UI/landing-page design, quick house repaint visualizations, kids’ T‑shirts.
  • Some see this as empowering small teams and non-artists; others fear erosion of design/illustration careers and a flood of low-effort “slop”.
  • Ethical debate runs from “free art for everyone” to “traitor to the human race”, with analogies to past technological shocks in creative industries.

Naming and product sprawl

  • The “Nano Banana” codename, embraced after social-media popularity, is widely seen as memorable but confusing.
  • Several users feel overwhelmed by Google’s growing AI product matrix (Gemini 3, Nano Banana, Antigravity, etc.), and worry about long‑term product stability given Google’s history.