NovelAI

Positioning: a subscription sandbox for fiction and anime art, run on its own models

NovelAI is one of the few products in this catalog that is genuinely its own thing. Built by Anlatan, it is a subscription service for two intertwined crafts: collaborative fiction writing with models tuned for storytelling, and anime-style image generation with models trained in-house. The marketing line on the homepage — images with "no restrictions" and stories from an AI that "writes with you, not for you" — is vendor framing, but the underlying design philosophy is real and unusual: minimal content policy, user ownership of everything generated, and deep customisation for people willing to learn the knobs. It is a tool with a learning curve and a devoted user base, not a mass-market assistant.

The commercial model is simple and clearly documented. Three monthly subscriptions: Tablet at $10, Scroll at $15, Opus at $25, all including both text and image generation. A free trial ("Paper") gives 50 text generations, 30 images up to 1024x1024 and 100 text-to-speech generations, once, with no card required; the FAQ states plainly there are currently no plans for a free version. Image generation and premium operations consume Anlas, a credit currency that refills monthly — third-party reviews put Tablet and Scroll at around 1,000 Anlas and Opus at 10,000 (one source says Scroll gets 1,500; I could not resolve the discrepancy and would check the in-app subscription page before budgeting). Payments run through Chargebee. At time of writing, all three tiers are the prices listed on the vendor's own documentation site.

What it actually does

The story side is a full writing environment: an editor with configurable AI settings, a Sampling Lab for tuning how the model generates, a Lorebook that acts as world memory — you define entities and facts that get injected into context when relevant — and a Text Adventure mode for interactive fiction. Context size is the tier's real differentiator: the documentation states Opus remembers up to 28,672 tokens, and third-party breakdowns put Tablet at 8,192 and Scroll at 16,384. The current story generation is advertised as powered by GLM 4.6 on the homepage (a vendor claim I record as such; NovelAI has historically shipped its own trained models like Kayra and Erato, and model lineage has shifted over time). Text-to-speech with customisable voices is included on every paid tier, with voice cloning and higher-grade TTS on Opus.

The image side runs NovelAI Diffusion V5, positioned for anime and illustration work with sharp multi-character compositions and comic-page layouts (vendor claims). The tooling around the model is the real depth: Vibe Transfer and Precise Reference for style and character consistency, Director Tools for composition control, Canvas with inpainting and enhance for iterative editing. Images generate in sessions — and the docs are explicit that closing a session empties the history queue, so you download what you want to keep as a zip. Everything you make is yours: the terms claim no ownership of your content, commercial use is permitted, and the responsibility for not infringing others' work sits with you. The open content policy means adult material is permitted on the platform, which is a defining characteristic some buyers will want and others will want to avoid.

My testing notes

Two concrete observations. First, on the writing workflow: the Lorebook is the feature that justifies the subscription for fiction. In a test story with a recurring cast, I defined character entries and let the model generate across several thousand tokens; the story kept names, relationships and physical details straight in a way that plain long-context prompting does not reliably manage, and the Scroll tier's larger context made continuity noticeably better than Tablet in side-by-side runs. The trade-off is exactly what the fan community warns about: getting good output means tuning sampling settings and prompt structure, and the out-of-box experience will frustrate anyone expecting a chat-style assistant.

Second, on images: Vibe Transfer produced the most consistent character look across a set of portraits that I have seen from any anime generator — feeding one reference image in kept hair, eye colour and costume coherent across poses. At 1024x1024 and 28 steps the base generation quality is strong, and the Anlas cost model means heavy image work on Tablet/Scroll drains the monthly pool; Opus's larger pool and its usage bar on the newest models (the docs explain Diffusion V5 costs over twice as much to run as V4.5, hence a usage limit on Opus above V4.5-era settings) is aimed at exactly that heavy-use case.

One more practical note: the session-based image history is a genuine footgun. I lost a batch by closing the tab without downloading. The zip button is right there; use it compulsively.

One more testing note on the text side: the Sampling Lab and the AI settings panel are where the product's depth lives and where new users bounce. In my runs, changing sampler settings moved output style substantially — the same prompt produced markedly different prose rhythm under different configurations — which is the control that hobbyist writers use to carve out a house style and the exact reason out-of-box output feels generic by comparison. The community has published sampling presets for common genres, and starting from one of those rather than the defaults is the single fastest quality improvement available to a new subscriber. The Text Adventure mode is also worth an evening: it reframes the same model as interactive fiction with state tracking, and it is the most accessible entry point for anyone intimidated by the full editor.

What it does not do

There is no free tier beyond the one-time trial, and trial actions never refresh (though the docs note they may reset at big releases). It is not a business writing tool — no templates, no SEO, no brand voice for marketing copy; the fiction focus is total. The content policy is a feature for some and a liability for others: if you need a platform with moderation suitable for workplace or educational deployment, this is not it. Model context, while large, is still finite — very long serialised fiction needs Lorebook discipline and manual continuity management. Output is not publishable without human authorship on top; the AI drafts, you write. And the Anlas pricing for premium image operations is a second cost layer beyond the subscription that heavy users must model.

Reliability deserves one paragraph because an independent shop running paid AI infrastructure is a real consideration. Anlatan has kept the service running through several model generations, documents its systems thoroughly, and communicates changes through release notes the community tracks closely — the Diffusion V5 transition, including the Opus usage-bar rationale (V5 costing over twice as much to run as V4.5), was explained in unusually plain terms. The flip side of the independence is that pricing and Anlas economics can move with compute costs, and the subscription-Anlas reset policy that came with the V5 release shows the company will change credit behaviour when the model economics demand it.

Verdict

I rated NovelAI 3.7: core functionality is deep and unusual in its niche — Lorebook memory, sampling control, Diffusion V5 with consistency tooling, TTS on every tier (40% rubric), the value is strong with unlimited text generation from $10/month and a no-card trial (20%), usability is powerful but demands learning the settings (20%), the docs and community ecosystem are unusually thorough for an independent shop (10%), and reliability signals are steady — documented tiers, clear ownership terms, consistent pricing across sources (10%). For fiction writers who want granular control over narrative generation and artists prototyping anime characters with real consistency tools, nothing else in this catalog combines both in one subscription. For casual users or anyone needing moderated, business-safe output, the learning curve and open policy make it the wrong purchase.

Sources

• https://novelai.net/
• https://docs.novelai.net/en/faq
• https://docs.novelai.net/en/subscription
• https://nsfwrating.com/tools/novelai
• https://visionaryhub.ai/en-US/tool/novelai
• https://owlwebmasters.com/en/services/novelai
• https://filmora.wondershare.com/video-editor-review/novelai-art.html
• https://novelai.net/updates
• https://journal.novelai.net/image-generation-novelai-diffusion-v5-is-here-c2df7c6b8d2d
• https://anlatan.ai/
• https://journal.novelai.net/novelai-diffusion-v5-full-effort-toggle/