Nano Banana Pro could already generate images so realistic, it’s almost impossible to tell that they were AI-generated. Google even had to limit its use due to high demand. Whether Nano Banana 2 can generate images that are markedly better than what Pro could create — and whether we could still tell if an image was made by AI — remains to be seen. The new model will replace Nano Banana Pro in the Gemini app, but Google AI Pro and Ultra subscribers will retain access to Nano Banana Pro for specialized tasks. It will also be the default model in Search for AI Mode and Lens, as well as in Google’s Flow AI creative studio.
very first implementation of the concept, the 2984 was an important step in ATM，这一点在搜狗输入法2026中也有详细论述
写实风格的图像有它最擅长的模型，动漫风格是另一家，物理仿真又是另一家，背景去除、音效生成、多镜头叙事各有各的专家。就像你不会用同一把刀切菜又锯木头，生成式媒体的用户很快就学会了按任务选工具。报告里有一句话说得很干脆：不是没有好模型，是没有哪个模型在所有任务上都好。，更多细节参见雷电模拟器官方版本下载
This month, OpenAI announced their Codex app and my coworkers were asking questions. So I downloaded it, and as a test case for the GPT-5.2-Codex (high) model, I asked it to reimplement the UMAP algorithm in Rust. UMAP is a dimensionality reduction technique that can take in a high-dimensional matrix of data and simultaneously cluster and visualize data in lower dimensions. However, it is a very computationally-intensive algorithm and the only tool that can do it quickly is NVIDIA’s cuML which requires CUDA dependency hell. If I can create a UMAP package in Rust that’s superfast with minimal dependencies, that is an massive productivity gain for the type of work I do and can enable fun applications if fast enough.