YMMV / NovelAI

• Broken Base: The art generator. On one side of the spectrum, supporters cite it as a great example of Genius Programming, being able to create artwork from lines of code and helping to immerse the stories they create, alongside producing visually stunning pieces (even if it takes several tries and careful wording). On the other hand, critics of it dislike people claiming generated pieces as their own work, and see art generated by neural networks as inherently conceptually too close to plagiarism for comfort due to the nature of it being trained on the work of existing artists.
• Fandom-Enraging Misconception: No, the image generator does not trace, or copy/paste into a collage, or keep a database of images that it looks at and steals from, or scour the Internet to swipe whatever you upload. That's simply not how the technology works, it's not even close to how it works.
• Fandom Rivalry:
  • With AI Dungeon 2, which, considering it was created by people who felt slighted by AI Dungeon's filters and perceived privacy violations, is only natural.
  • With Spellbound Storytelling, a character-based story generator, especially after it took out ads. Supporters of this service would rather it replace NovelAI entirely due to the latter's slow development and less powerful model, which is a death sentence in the quickly-developing world of AI text generation technology. It is also similarly encrypted in usage, a quality people also liked from NovelAI. NovelAI supporters, including embarrassingly, the NovelAI community manager, Aini, reacted badly against this, bringing up Spellbound's vitriol against NovelAI when Spellbound was being promoted.
  • Within the userbase themselves, the classic Old Guard Versus New Blood, especially concerning text generator vs. image generator. The former was the reason most users subscribe, and with the name "Novel" AI, it was what appealed to them in the first place. The latter was developed later and it was at first praised for its Genius Programming and becoming an easy money-maker due to the mass appeal for image generators, with its currency Anlas being able to be bought separately if the user does not have the infinite generation attempts from the Opus subscription tier. However, the split became clear when it was seen that there was more effort put into the image generator going by its more frequent update logs that the oldschool users only interested in storywriting felt alienated and began to stop subscribing. The absolute nadir was the v3 update to the furry model because it was yet another image generation update when the latest text generator model, Kayra, was released about half a year prior. Text generator fans called it quits (with plenty of support to do this) and even blamed the image generator fans, while the latter defended by stating that this is the first major update for that specific model that finally allowed it to achieve the same Genius Programming as the latest anime model and called out the impatient users.
• Germans Love David Hasselhoff: For better or for worse, this model, especially the art generator, is very popular in Japan, especially for the model trained on anime images. The text generator is no slouch there either, with one of the AI text models being specifically trained on Japanese-language works for this reason (though that is a work in progress), as well as a Japanese localization for the site and a new tokenizer with special support for Japanese being introduced in early 2023.
• Memetic Mutation: Is textgen/imagegen dead?^ExplanationNovelAI's notoriously slow development for an LLM-based AI service combined with how one feature's tendency to have multiple update logs in a row (usually the image generator) caused recurring concerns from users asking if one service (usually the text generator) has been abandoned. This was enough to become a community in-joke, even to the point when the text generator finally got the update it needed, the image generator became the "concern" topic instead. With Aetherroom being announced and in development, either generator can take the place, or even both.
• Once Original, Now Common: When NovelAI was made as a reaction to the censorship in AI Dungeon 2, people began looking forward to it, especially when the latter was hitting rock bottom in model quality no matter which model. People also began to use its API to complement it with other AI-based apps/games, which is helped by how relatively more powerful it was for something developed in-house. Everything being encrypted also won users over because of, again, scandals related to AI Dungeon 2, this one where other people can read even the users' unpublished stories. However, more AI services sprang up with the same or even better quality for cheaper or even free, and even people are more appealed by ChatGPT's instruct-style interaction (which NovelAI eventually applied but not so well). A few, like Spellbound Storytelling, even encrypted requests and stories, do not have the same limitations as OpenAI-based services, and do not even require users to sign in to write stories. With how slow NovelAI's development is in both text and image department, long-time users have started complaining or even jumped ship while curious newcomers are alienated by the Difficult, but Awesome tweaking just to get a good narrative or image when it can be easily done elsewhere. Heck, even AI Dungeon 2 itself has caught up as of 2024, even with censorship.
• Scrappy Mechanic:
  • Even though the image generator uses a modified version of the Stable Diffusion model, the base training data module for Stable Diffusion is not included, instead only having modules trained on images from Danbooru (for the anime modules) and e621 (for the furry module). While the reason for this is the extensive and highly-specific tagging systems on those sites making training much easier, as well as avoiding the disputed legal and moral ramifications of some of the uses people have for the base Stable Diffusion module, it still causes no end of frustration for people who want to create images that are neither anime nor furry.
  • The difficulty of getting characters in images to actually hold objects such as weapons in their hands — let alone correctly, with things such as characters holding swords by the blade or guns by the barrel being a common sight — is notorious among the community.
  • The AI tends to be overzealous about placing dinkuses (***), which indicate chapter breaks, and asterisms (⁂), which indicate story breaks, both of which lead to scenes and stories being Cut Short if not manually erased.
  • Text in the Memory and Lorebook tends to leak into the main story verbatim if that is not specifically prepped against.
• Spiritual Successor: Novel AI is largely staffed by people unsatisfied with changes that AI Dungeon underwent in 2021.