nai-atelier

NAI Atelier is a NovelAI creation workshop that runs on a personal computer. It is a secondary development based on kirafishy/NaiPromptManager, independently maintained, and licensed under MIT. It is not merely a prompt input box, but a complete workflow built around long-term personal use: collecting artists and characters, combining prompts, calling NovelAI, saving original images and parameters, organizing history and inspiration, and continuing to use it between computer and phone. Scope and Prerequisites The project is explicitly intended only for NovelAI Opus tier subscribers. Cost estimates, free quota determinations, and Opus quota displays are all designed for Opus; other tiers and third-party relay services are not adapted. Running it requires Node.js 24.x (or 22.19.0 or above) and npm, as well as an API Key starting with pst-. For first use, start it via git clone, npm install, and npm run dev:local. Once the local gateway and lexicon services are ready, access localhost:3000. No deployment to Cloudflare is required. Core Functional Modules Style strings are used to save reusable art styles, modules, parameters, and covers. Prompts and modules are decoupled, supporting independent positive and negative terms and coordinate positioning for multiple characters. The image generation lab supports four modes: text-to-image, image-to-image, inpainting, and outpainting. It provides size and cost estimation, step locking, transparent channel generation, PNG/JSON parameter import, and SSE streaming process image previews. Generation history persists original images and parameters locally to D1 and R2, supporting filtering, pagination, batch cleanup, and one-click sending back to the lab. The inspiration library serves as a personal knowledge base, supporting tags, notes, inspiration boards, and ratings, and can be derived into style strings, character references, or permanent Vibes. Resource Libraries The project provides a GitHub download entry for the ffdkj Chinese-English Tag database. It does not include a built-in database or pre-generated lexicon. Users need to download it in settings or run npm run update:tags to generate local shards, with a scale of about 320,000 entries. The artist library can generate about 151,000 artist entries, supporting search, sorting, favorites, random draws, and generating baseline preview images with a unified prompt. The character library distinguishes between official Tags and custom restored characters. AITag directly connects to aitag.win to search NovelAI works and reverse-lookup prompts and parameters, supporting import to the lab, saving as style strings, or adding to the inspiration library. Danbooru connects to an all-ages image library for materials and Tag references. Pixiv provides rankings and tag browsing through official OAuth login. Image reverse inference uses a local WD Tagger, with an initial download of about 379 MB for the model, and no need to upload to the cloud. Project Agent and Assets The project Agent is based on the lightweight pi.dev core. It can read local materials, observe historical original images, and call workshop tools. It supports multiple sessions, multimodal visual collaboration, and draft-box-style transactional submission; operations involving deletion or consuming Anlas require frontend confirmation. Permanent Vibe Transfer encodes reference images into irreversible encodings. Each new encoding consumes 2 Anlas, and it supports combination, import/export of .naiv4vibe, and deduplication. Precise Reference supports three types: character, art style, and character plus art style. Each reference image and generated image additionally consumes 5 Anlas. Phone and Interoperability Phones access port 3000 of the computer through the home LAN, using four-digit PIN authentication. The local machine is password-free, and after verification, no input is required for 30 days. The mobile end is rebuilt for touch, with main controls meeting 44×44px, persistent bottom navigation, and dedicated layouts and cache settings for each module. The project includes a built-in connector for st-chatu8, which can be installed into SillyTavern with one click, synchronizing style strings, Vibes, and generation history in both directions. Historical images are indexed by path to avoid duplicate disk usage. Cost and Privacy The project distinguishes between Anlas points and Opus quotas, recording local budgets and usage by Key, and displaying estimates and confirmation before generation. The multi-key vault supports alias management and switching, but keys are currently stored in plaintext JSON in local D1; interface masking does not equal encryption. Image safety mode provides a triple mask of blur, darkening, and desaturation, supporting temporary peeking and out-of-focus self-locking. Data and Backup Original images, thumbnails, encodings, and recovery records are located in the local-data persistent area. Clearing browser cache will not delete records, but long-term retention still depends on backups. If the final write of a Vibe encoding fails, it will first be saved to local-data/vibe-recovery and automatically entered into the database after restart, avoiding duplicate payment.