Overview
NovelAI is a subscription service built around two connected tools: a story-writing editor powered by a large language model, and an image generator built on NovelAI's own diffusion models rather than a widely shared open-weight base. Both are accessed through the same account and subscription, and both are marketed explicitly around creative freedom, including for adult-oriented material, rather than being general-purpose assistants with adult use as an incidental side effect. The writing tool is an editor rather than a chat interface: users write alongside the AI, which continues, expands, or rewrites story text based on what's already been written, and can be pointed at first-, second-, or third-person perspective.
Text models and story steering tools
NovelAI has offered a series of text models over time, including Kayra, an in-house 13-billion-parameter model trained on roughly 1.6 trillion tokens, and Erato, built on a Llama 3 70B base with additional training on NovelAI's own pretraining dataset; its current top-tier text model is a fine-tune of an open-source Mixture-of-Experts model with 355 billion parameters. Story steering tools include Lorebooks (persistent reference entries the model can draw on for characters, settings, or world details), a memory/notes system, reusable prompt "modules," and adjustable generation parameters like temperature and sampling settings. Context window sizes vary by model and subscription tier, with documentation citing up to roughly 28,000 tokens of usable context on newer models, versus 8,192 tokens on older models like Kayra and Erato.
The Lorebook system
A Lorebook entry has a title (for the author's own organization, invisible to the model), body text, and one or more activation keys. When a key appears in the recent story text, the entry's content is inserted into context automatically; keys can be plain text, regex patterns, or combinations requiring multiple keys at once, and an "Always On" toggle can bypass key-matching entirely for an entry that should always be present. A Placement tab controls exactly where a triggered entry gets inserted, how large a token budget it can claim, and what gets trimmed if space runs short. Entries can be grouped into categories that share default settings, and a lorebook generator tool can draft new entries from a prompt.
Memory and Author's Note
Separate from Lorebook entries, NovelAI provides a Memory field for information the model should treat as always true regardless of how far back in the story it was established, and an Author's Note field for injecting steering instructions (tone, style, focus) at a fixed point near the end of context. A Context Viewer shows, with color-coded sections, exactly what text is being sent to the model on a given generation and how many tokens each source (story, Memory, Lorebook entries, Author's Note) is consuming. This can help diagnose why a Lorebook entry didn't fire or why older story text got trimmed. The editor offers two base modes on the same underlying tools: Storyteller, aimed at conventional prose writing, and Text Adventure, which frames input and output more like an interactive-fiction game with its own default system prompt. A built-in Text-to-Speech option can read generated story text aloud.
Image generation and Vibe Transfer
NovelAI's image generator is built on its own NovelAI Diffusion model family, which has moved from an earlier version built on latent diffusion techniques adapted from SDXL (NovelAI Diffusion Anime V3 and Furry V3, at a base resolution of 832x1216) to newer V4 and V4.5 models trained from scratch rather than fine-tuned from a public base model. The tool supports text-to-image generation, image-to-image (with separate Strength and Noise sliders controlling how far a generation departs from the source image), inpainting, and upscaling. Its distinguishing tool is Vibe Transfer: a user uploads a reference image, and two sliders control the result: Reference Strength (roughly 0.01–1.0, default 0.6) sets how strongly the output emulates the reference's visual style, while a separate Information Extracted value controls how much fine detail versus broad composition is pulled from that reference; it can be combined with an ordinary text prompt rather than replacing one. Newer image models support multiple characters in a single generation and natural-language prompting alongside traditional tag-based prompting, plus reusable Undesired Content presets that function like a saved negative prompt. Batch generation is capped by image size: up to 6 images at once at smaller sizes, 4 at normal/large sizes.
Content policy and adult-content positioning
NovelAI's Terms of Service require that users be 18 years or older to access the service. Consistent with that age requirement, NovelAI is widely known for allowing adult and NSFW content in both its text and image tools, without the blanket filtering that mainstream writing and image platforms typically apply; this is a defining, publicly marketed characteristic of the platform rather than a loophole. NovelAI's documentation does not publish a detailed content-moderation policy beyond the 18+ requirement in the same way some other platforms do.
Access model and pricing
NovelAI operates on tiered monthly subscriptions (published on its site as Tablet, Scroll, and Opus), with higher tiers unlocking larger text context windows, higher included image resolutions, and a larger monthly allotment of Anlas, an in-platform currency used to pay for image generations beyond a plan's included allowance. Anlas included with a subscription refill each cycle, while separately purchased Anlas do not expire. Opus, the top tier, additionally allows single-image generations to bypass the Anlas cost entirely when they meet specific published conditions: no base image used, "Normal" size or smaller (up to 1024x1024), and 28 or fewer generation steps, anything outside those conditions draws from the plan's monthly Anlas allotment instead. A free trial tier gives new, verified accounts a fixed number of free text generations, text-to-speech generations, and images before a subscription is required.
Limitations
Because Vibe Transfer and image-to-image both depend on an uploaded reference or source image, results are sensitive to the strength/noise values chosen and generally require some trial and error to dial in: NovelAI's own documentation frames these as tunable parameters rather than a one-click style match. Free image generation is also narrowly conditioned (Opus tier only, single images, capped size and steps), so most image work outside that window draws down Anlas regardless of subscription tier.
Important distinctions
NovelAI differs from character-chat platforms like Character.AI or JanitorAI in that its core writing product is a collaborative story editor, not a persistent character-persona chatbot, even though its Lorebook system serves a similar purpose to character memory in companion apps. On the image side, it differs from tools built on shared open-weight checkpoints, such as those hosted on Civitai or Hugging Face, because its current-generation image models are proprietary and trained in-house rather than being publicly downloadable weights.
Related glossary concepts
Related database entries
Sources & further reading
- NovelAI — Official site
- NovelAI Documentation — FAQ
- NovelAI Documentation — Image generation
- NovelAI Documentation — Image models
- NovelAI Documentation — Text models
- NovelAI — Terms of Service
- NovelAI Documentation — Lorebook
- NovelAI Documentation — Advanced Settings (Context Viewer, Phrase Bias)
- NovelAI Documentation — Vibe Transfer
- NovelAI Documentation — Subscriptions