Nano Banana 2 Lite

Original Article ↗ Hacker News Discussion ↗

Google’s new “Nano Banana 2 Lite” (a fast, lower‑cost Gemini image model) is praised for sharply reduced latency and decent quality, making it attractive for bulk or interactive use cases like kids’ story apps, while still trailing more advanced models such as ChatGPT Image 2 on nuanced prompts and aesthetics. Commenters weigh trade‑offs between speed, price, instruction‑following and censorship across competing image systems, and note rough edges in Google’s product access, pricing and resource limits. A major thread questions the ethics of AI-generated interior photos in real estate and rentals, with many arguing that idealized or physically impossible images amount to fraud and should trigger stronger regulation or technical countermeasures.

Model performance & comparisons

  • Many feel ChatGPT Image 2 is notably stronger than Google’s models in aesthetics, detail, and especially nuanced prompts; some call it “insane” and wonder why it wasn’t in Google’s comparison chart.
  • However, its latency is much worse (often ~2 minutes at 1024×1024), making it less suitable for fast workflows.
  • Nano Banana 2 Lite (NB2L) is described as a distilled, faster, cheaper version of NB2, with worse performance on highly nuanced prompts but better text rendering than NB1.
  • Some users prefer Google’s image models overall for their workflows; others say Gemini is still “behind,” especially compared to OpenAI and certain newer competitors (e.g., Grok, Krea2, Ideogram) on quality or benchmarks.
  • Public image benchmarks and ELO leaderboards are widely criticized as noisy, gamed, and biased toward aesthetics over instruction-following.

Latency, cost & practical use cases

  • NB2L’s key selling point is speed (often a few seconds vs ~30s for NB2, vs much longer for ChatGPT Image 2) and slightly lower price than NB1.
  • Some find the price still high for personal use but acceptable for enterprise and API-heavy workflows.
  • Use cases mentioned: bulk/report imagery, blog illustrations, fast prototyping, kid storybooks with likeness, photo restoration, and bathroom remodel mockups.
  • Users note different priorities: high-end art workflows tolerate cost/latency; embedded or “throwaway” images need cheap and fast.

Capabilities & limitations

  • NB2L supports aspect ratios via certain APIs (e.g., Vertex), contradicting an early claim that it does not.
  • Editing behavior is reported as improved vs previous Gemini image models but still degrades over multiple edits.
  • Text in images often comes out garbled and may appear unprompted, though negative prompts can sometimes suppress it.
  • Some are frustrated with frequent safety or refusal messages on news-related prompts and children-related content.

Access, tooling & UX

  • Complaints about Google’s fragmented product tiers: differences between Gemini app, AI Studio, Workspace, and Google One access; some need multiple paid accounts.
  • Workarounds include third‑party tooling (e.g., generic API clients, OpenRouter, other frontends) to unify model access.
  • Google’s infrastructure sometimes returns RESOURCE_EXHAUSTED errors under parallel load.

Ethical and legal concerns (especially real estate)

  • Large subthread on AI-generated or heavily edited real estate photos: many describe them as deceptive or outright fraudulent.
  • Examples include impossible furniture layouts, fake fixtures, altered room dimensions, unrealistic lighting, and changed window views.
  • Some argue this should clearly fall under false advertising/fraud and be illegal or better enforced; others note similar misrepresentation predates AI (wide-angle lenses, Photoshop).
  • There is discussion of emerging regulations (e.g., disclosure requirements), MLS rules against altering property condition, and weak consumer protection enforcement.
  • Ideas floated: lawsuits, stricter liability (force landlords/sellers to match advertised features), or even browser plugins/AI tools that “de-fake” listing images.
  • Broader concern that AI tools lower the cost of fraud and add “economic friction” without creating real value.

Watermarking & provenance

  • Google says images carry invisible SynthID watermarks.
  • Some see this as necessary to avoid an unmarked flood of AI images; others dislike any mandatory watermarking of their artwork, viewing “AI risk” messaging as overblown and power-consolidating.

Prompting & marketing

  • Example prompts on the product page are derided as obviously machine- or copywriter-generated and unrealistic for actual users, who tend to use concise, targeted prompts instead.