NovelAI: the team that has set the bar for anime generation for five years
Every generation of nai-diffusion changed how the community draws anime with AI: from the leaked model that SD 1.5 grew on to V5 with its 32-channel VAE. Below is an honest timeline and what you can try for free in our bot.
From the leak to V5 — six turning points
- 2021
Anlatan launches NovelAI
First as an AI co-writer for stories: their own text models on their own servers, a focus on privacy and full user control over the output. The Discord server opened on 28 April 2021 and now has over 56,000 members. A small independent team in the habit of training models itself rather than renting someone else’s.
textDiscord 56K - 2022
NovelAI Diffusion and that leak
In October 2022 NovelAI shipped its first anime model, built on Stable Diffusion and fine-tuned on Danbooru: tags instead of sentences, quality tags, hypernetworks. Within days the weights leaked — and the entire SD 1.5 era grew out of them: Anything V3, hundreds of merges, the familiar WebUI presets. The “prompt = Danbooru tags” standard that every anime model uses today comes from here.
SD 1.5Danbooru tagsAnything V3 - 2023
NAI Diffusion Anime V3: SDXL unlocked
The first anime model on SDXL, and the one that showed the community what the architecture could do: clean lines, 832×1216, stable anatomy. The big discovery was artist tags: as109, ciloranko and thousands of other web artists could be reproduced by name, and mixing them produced new styles. The “one prompt × ten artists, one seed” culture started with V3. Furry V3 and inpainting followed.
SDXLartist tagsas109ciloranko - 2024
Vibe Transfer and Director Tools
Vibe Transfer made it possible to copy the style, palette and mood of any image without a single word in the prompt — several references at once, with a dial for how much composition versus style to take. Director Tools added background removal, line art, colorization and changing a character’s emotion. NovelAI stopped being just a “tag generator”.
Vibe TransferDirector Tools - 2025
V4 and V4.5: an architecture of their own
NovelAI moved away from SDXL: V4 is a fully in-house architecture with a T5-based text encoder, natural language mixed with tags, up to six characters with positions on the canvas and the first text rendering. V4.5 added a background dataset, the masterpiece tag and such a quality jump that Anlatan “considered calling it V5”. Precise Reference with its Character Reference mode is like a LoRA trained on the fly: a character from a single reference even if it is not in the dataset; the Style mode does the same for style.
T5Character ReferencePrecise Referencepositions - 2026
V5: the current state of the art
Still 1 MP on output, but the model is 2.5× the size of V4.5 and was trained in-house over 268,000 B200 GPU-hours. The headline is a 32-channel VAE instead of 16 channels: twice the detail where things used to melt — eyes, hands, jewellery and small text on full-body characters. The text encoder changed too: instead of T5, a model from the Qwen3 family (per community findings; Anlatan does not disclose the architecture), the prompt budget grew from ~512 to ~1,471 tokens, English and Japanese are officially supported, and in our tests V5 understands Russian without translation. Up to 22 characters with free positioning, comic pages in a single generation, text up to 750 characters in English, Japanese and Chinese, transparent backgrounds. 12.6 million images in the first 24 hours after release.
32-ch VAEQwen322 characterscomicstext
Six things NovelAI did first
Vibe Transfer
Style and mood from any image without words: upload a reference and get its “vibe” in your own scene. A dial controls how much composition versus manner to take.
Character Reference
A character from a single reference: a recognisable face, hairstyle and outfit in new poses and scenes, even if that character is not in the training data. Works like a LoRA you never have to train.
Character positioning
A separate prompt for each character and a point on the canvas where they stand. V4: up to six on a 5×5 grid; V5: up to 22 with free coordinates and no feature bleed.
Text on the image
Signs, comic speech bubbles, T-shirt prints. V4 — English only, 118 characters; V5 — English, Japanese and Chinese up to 750 characters, just put the phrase in quotes.
Comics
Describe the page layout in plain language or place the characters on the canvas — V5 returns a finished multi-panel page in one generation. Earlier models could only do simple strips.
Transparency and Enhance
The V5 VAE natively supports an alpha channel: sprites and translucent effects without cut-outs. Enhance “Max” upscales and sharpens in one click on the new upscaler.
What it looks like in practice
Anlatan’s V5 teasers and user artworks from @novelai_bot on V4.5.
NovelAI directly or through @novelai_bot
We do not replace NovelAI — we make it reachable from a chat, paid in roubles. An honest comparison:
| novelai.net | @novelai_bot / VK | |
|---|---|---|
| Subscription | $10–25 per month, international card | From 150 ₽ per month: Russian card, SBP, Telegram Stars |
| Free | 30 images once (trial) | 20 → 10 V4.5 generations every day, no card |
| V5 | Opus $25 with a usage “battery”; Tablet/Scroll pay per Anlas | On every paid plan with a daily limit |
| Access | Registration, website, payment from Russia is difficult | Telegram or VK, nothing else needed, works without a VPN |
| Tools | Full UI: Vibe Transfer, Precise Reference, inpainting, Director Tools | The bot’s prompt tools: !style, positions !a1–!e5, the /tt tagger, i2i. No Vibe Transfer or Precise Reference yet |
| Prompt language | English, Japanese | Russian too: V5 understands it without translation, for V4.5 the bot translates |
Anlatan does not just ship models — it invented the language the community uses to talk to anime models: Danbooru tags, quality tags, artist tags, character positions. All of it later showed up in Anima, Illustrious, Pony and other open models. That is why NovelAI sits next to them in @aniworld_bot rather than instead of them: compare for yourself on one prompt.
Sources: journal.novelai.net, docs.novelai.net, novelai.net/v5, the novelai.net/updates changelog, the @aniworldai channel.
Questions about NovelAI
What is nai-diffusion?
It is the API name of NovelAI’s image generation models: nai-diffusion-3 (Anime V3 on SDXL), nai-diffusion-4-full, nai-diffusion-4-5-full, nai-diffusion-5-full and nai-diffusion-5-curated. The company is Anlatan, the product is NovelAI.
How is V5 different from V4.5?
The model is 2.5× larger, a 32-channel VAE replaces the 16-channel one (visible in eyes, hands and small details), a new text encoder with a ~1,471-token budget instead of ~512, up to 22 characters instead of 6, free positioning, comic pages, text in three languages up to 750 characters and transparent backgrounds. The resolution stays at 1 MP. Vibe Transfer and Precise Reference for V5 are promised later.
Does NovelAI understand Russian?
Officially V5 supports English and Japanese. We ran an A/B test in production: three Russian prompts with one seed, raw Russian versus a translation — all six artworks contained every element of the prompt, with differences within seed noise. So the bot sends Russian to V5 untranslated. For V4.5 the bot translates the prompt itself.
Can I try NovelAI for free?
Yes. In @novelai_bot on Telegram and VK newcomers get 20 NAI Diffusion V4.5 generations per day, then 10 per day, no card. V5 is available on paid plans from 150 ₽ per month with a daily limit. On novelai.net itself only 30 images are free, once.
Try NovelAI in your language
Open @novelai_bot, type /nai and comma-separated tags. 20 free generations on the first day, no card, no VPN.