- Author
- Posted by minimaxir, with replies from many commenters
- Published
- 13 November 2025
- Kind
- Hacker News discussion thread about a long blog post with generated images
- Original
- news.ycombinator.com/item?id=45917875
- Subject
- AI put to use
What it is
The thread discusses a blog post by Max Woolf, posted to Hacker News by minimaxir. The post tests the model with complex prompts, JSON character sheets and edits. Commenters share their experience, challenge claims and branch into other topics.
What it says
- Max Woolf's linked post says Nano Banana follows complex prompts very closely.
- The linked post says Nano Banana is terrible at style transfer. New images in a named style work.
- Posting as minimaxir, Woolf says one prompt took 28 iterations.
- Several commenters say Nano Banana still makes random edits, changes scale or adds objects.
- Commenters note a left-right ambiguity in the skull-pancake edit. Woolf concedes the miss.
- One commenter storyboards with seven prompt layers and red boxes marking edit regions.
- A long subthread argues over taste, gatekeeping and artists' objections to AI.
What evidence it gives
The post shows prompts with generated images and says notebooks for each image are released. Commenters mostly give personal anecdotes, a few shared chats and links to their own tools. No commenter ran a controlled comparison.
Interests and limits
The post is one author's hands-on testing of one model. The author says in the post to be seeking a job in AI. Several commenters promote their own products. Hacker News comments are unverified, and some posts are off topic. The images in the linked article and in comment links were not viewed.