A 23k-star scrapbook of Gemini-2.5-flash-image experiments

What it does This repository is a curated gallery of images and prompts generated by Nano Banana and Nano Banana Pro, a model built on Gemini-2.5-flash-image. The maintainers gather standout examples from Twitter, Xiaohongshu, and other platforms—ranging from Ukiyo-e flash cards and toy disassembly diagrams to fake Polaroids and city weather cards—and organize them into a browsable catalog. It also hosts Nano-consistent-150k, an open dataset of over 150,000 samples focused on keeping the same character identity across dozens of different editing tasks.

The interesting bit The dataset is the real payload. While the gallery is entertaining, Nano-consistent-150k is built so a single portrait can survive more than 35 different edits—hairstyle swaps, outfit changes, age progression—without turning into a different person. That interleaved structure is meant to train unified models that actually remember who they are editing.

Key highlights

• 52+ “Pro” case studies and dozens of standard cases, each with input/output pairs and social-media attribution
• Nano-consistent-150k: >150k samples for identity-consistent image editing, released on Hugging Face
• Multi-modal prompts: examples span infographics, 3D isometric rooms, AR overlays, manga storyboards, and math-reasoning visualizations
• CC BY 4.0 license on the repo; multi-language READMEs (Chinese, English, Japanese, Korean, Spanish, Turkish)
• Sourced from live community posts on Twitter/X and Xiaohongshu, so the collection reflects real user experiments rather than lab benchmarks

Caveats

• This is a gallery and dataset repo, not a model weights or inference codebase—there is no code to run Nano Banana here
• Technical details on how Nano Banana differs from base Gemini-2.5-flash-image are absent from the README
• The primary language is Chinese; other translations exist but the core content is Sino-centric

Verdict Worth a bookmark if you are building image-editing datasets, hunting for prompt engineering inspiration, or researching identity consistency in diffusion models. Skip it if you are looking for training code, model weights, or a local inference pipeline.