Nano Banana image examples

Original Article ↗ Hacker News Discussion ↗

Perceived Capabilities and Progress

  • Many commenters are struck by how far image models have come: consistent characters, complex compositions, localized edits, and convincing photo-style results from simple prompts.
  • Nano Banana is identified as Google’s Gemini 2.5 Flash with native image output, tuned primarily for editing; praised as fast, cheap (~$0.04/img), and near state-of-the-art.
  • Benchmarks show it leading or near the top for image editing, but strong competitors (Seedream 4, Flux/Flux Kontext, Qwen Edit, GPT‑image‑1) sometimes outperform it, especially in open-weight or local settings.

Reliability, Adherence, and Cherry-Picking

  • Multiple users report that the showcased examples are heavily cherry‑picked, often requiring many “rolls” to get one good result.
  • A common failure mode: the model ignores requested edits and returns nearly the same image, or miscarries details like poses, aspect ratios, and object placement.
  • Prompt engineering strongly affects quality; structured, LLM-style prompts and “award-winning/DSLR” style phrases, plus long-context JSON/HTML, improve adherence—but results remain non-deterministic and fragile.

Impact on Artists and Work

  • Debate over whether professionals should “learn the tool or change careers.” Some argue prices will collapse but skilled artists using AI will still outperform amateurs, similar to digital cameras and photography.
  • Others note current unreliability means AI can’t fully replace designers or illustrators yet, but it does remove a huge amount of “pixel-pushing” work.

Safety, NSFW Bias, and Workplace Concerns

  • Early examples included a sexualized anime panty-shot; this was quickly removed after complaints about NSFW content and workplace appropriateness.
  • Ongoing tension between calls for uncensored models and concerns about harassment, cultural norms, and the visible bias toward young, sexualized women in many demos.

Technical Gaps and Limitations

  • Text and diagrams are often wrong: anatomy labels, building names/dates, UI text, and map/topography interpretations look plausible but are factually incorrect.
  • Struggles with clocks, wireframes, precise camera specs, transparent backgrounds (fake checkerboards), and some composition tasks (multi-angle product shots, real-photo integration).
  • Safety filters frequently block benign edits of real people, frustrating users.

Misuse, Trust, and Authenticity

  • Many worry about an oncoming wave of convincing deepfakes, fraud, and disinformation, arguing we’re approaching a point where online imagery is broadly untrustworthy.
  • There is discussion of cryptographic provenance standards (e.g., C2PA) and signed-camera ideas, but skepticism that these can fully solve authenticity or “photo of an AI scene” problems.