Glossary

The AI persona of Nephele Workshop. She is the operator of the Workshop and helps you complete tasks through conversation.

A desktop workstation for artists built with PySide6 + QML, offering AI chat, browser automation, and batch Pipeline processing.

The Agent is Nephele's core capability. Unlike ordinary chatbots, the Agent actually operates your computer: opening browsers to search images, organizing Eagle assets, batch-packing works, and running Python scripts. It understands intent via the Skill system and invokes 60+ tools through the ToolRegistry. Tools load on demand — each turn only sends the model the tools relevant to your request and fetches the rest dynamically when needed, saving a large amount of context.

The local zero-latency intent parser. Clear commands like 'open Photoshop' or 'open the Downloads folder' skip the AI entirely — the Rule Engine executes them directly, near-instantly, at no cost.

The full Agent is driven by a model tier you pick in the main input box (like choosing a model in ChatGPT): Axioma Zephyr (fast, billed at half rate) by default, or Axioma Tempest (full-power, full rate). Both have a 1M-token context, both run a ReAct loop (reason → call tool → observe → reason again) with up to 20 tool calls per message, and both are billed in Nepheline. A local 0ms rule engine still handles explicit commands, and a separate free '?' Guide channel absorbs light 'how do I use this' questions.

The Guide channel surfaces as the '?' popover in the top-right toolbar: a deterministic chip flow (purely local, no LLM, copy sourced from flows.yaml) plus free-form input (streamed answers aware of which screen you're on). When it can't answer or you want to go deeper, it 'escalates' to the full Agent (Tempest tier, asking you to confirm first) carrying your exact words. It's biased toward passing through and fails open — it will never block your input. It never generates UI operations, only returns structured whitelisted actions.

Nephele's cloud AI inference service. The full Agent offers two tiers: Axioma Zephyr (fast, half rate) / Axioma Tempest (full-power, full rate), both with a 1M-token context; the free '?' Guide channel runs on a separate lightweight lite model.

Skills are Nephele's secret weapon. The same prompt to different AIs yields wildly different results. Nephele's edge isn't a smarter model — it's Skills. A Skill is a plain text file describing in natural language 'when the user says X, here's what you should do'. It's not a plugin, needs no installation, and requires no code. Skills load from three sources: cloud push, user custom, and built-in.

Atomic function units the Agent can invoke, such as eagle_search, pack_image, evaluate_image — 60+ in total, loaded dynamically on demand.

Nephele's tool registry singleton managing all Agent-available tools, supporting dynamic registration and remote disabling.

The Agent's core working pattern (Reasoning + Acting): reason, call a tool, observe the result, reason about the next step, and repeat until done. Up to 20 tool calls per message.

Automated processing pipeline that chains multiple tools in sequence, e.g. 'tag → evaluate → pack → certify' in one click.

Fatigue points — the test-period daily free quota, refreshing every day at midnight (Beijing time) for trying cloud features. The system always deducts stamina first, touching Nepheline only once it runs out.

Nepheline is Nephele's only paid currency, used for all cloud AI inference (the full Agent, Creative Review, StyleDNA, etc.). An annual update pass / license purchase bundles 50,000 Nepheline; purchased and bundled Nepheline share one pool, never expire, and have no cap. When the license lapses, only cloud AI features are disabled — your Nepheline isn't wiped and is restored on renewal. Deduction order is fixed: daily stamina → bundled Nepheline → purchased balance. Running out only pauses cloud AI; local features are unaffected. Light questions absorbed by the Guide channel are never billed.

StyleDNA scores your art via a cloud vision model across 5 bipolar dimensions (line/palette, realistic/stylized, minimal/maximal, vivid/unsaturated, narrative/atmospheric), encoded into a 5-letter code mapped to one of 32 style types — a personalized artist identity card.

Dynamically adjusts evaluation dimensions by artwork type, giving a ten-tier S→E grade (S / S- / A+ / A / B+ / B / C+ / C / D / E) and improvement suggestions across composition, color, technique, and more.

One-click export of watermarked high-res PNGs. Supports full-image watermarking and blind watermark steganography. PSD files can be processed directly.

Auto-generates platform-adapted copy and tags for multiple social platforms. It does not upload — it only writes the copy.

256-bit steganographic protection embedding copyright info into the image's frequency domain — invisible to the eye but extractable for verification. Built on the open-source blind-watermark algorithm (DWT-DCT-SVD).

Digital Certification gives a work an independently verifiable 'birth certificate'. It hashes the file, requests an RFC 3161-compliant trusted timestamp from a third-party Time Stamping Authority (TSA), and generates a QR-coded PDF forensic report. In batch certification all files are chain-linked in a Merkle tree, so tampering with any one breaks the whole tree. Verification relies only on the public timestamp protocol and CA certificates, not Nephele's servers — even if Nephele shuts down, the evidence you've issued remains verifiable by courts and third parties.

On detecting infringement: auto-screenshots the full webpage, saves source code, extracts images, captures TLS certs — all timestamped into a tamper-proof evidence package.

The self-contained evidence file produced by Evidence Capture, bundling page screenshots, HTML source, the TLS certificate chain, and network logs — all timestamped. Independently verifiable at verify.arisfusion.com without relying on Nephele's servers.

Upload an image to search its source across engines in parallel: SauceNAO, IqDB, Google Cloud Vision, and Baidu Image Search, plus local Eagle-library matching and exact Booru MD5/hash lookup.

Finds visually similar images within your local Eagle library. Uses the EVA02 tagger embeddings (1024-dim, produced for free during tagging) for cosine similarity, blended with tag-overlap ranking — fully local, no network.

Compresses image content into a fingerprint robust to resizing, compression, and minor edits (pHash / dHash / aHash), computed locally via the open-source imagehash library. Nephele uses it for similarity matching and tracing stolen images.

Get reference images via Agent chat or manual search across Pixiv, Konachan, Pinterest, Huaban, ArtStation, Wikimedia Commons, and your local Eagle library (Eagle-only by default; network sources are opt-in).

Eagle is an asset management app commonly used by artists and designers. Think of it as an enhanced folder specifically for collecting and finding images. Nephele's relationship with Eagle is filesystem-level compatibility, not a runtime dependency: Nephele reads and writes Eagle's .library/ filesystem and metadata.json directly, regardless of whether Eagle itself is running.

Nephele's built-in file browser docked at the bottom of the window, unifying local files and Eagle asset browsing.

Library indexing runs models locally via ONNX Runtime in a single pass: by default the PixAI Tagger v0.9 (EVA02 architecture, ~1.2GB, character vocabulary through 2025-01) recognizes image content and applies tags, harvesting a 1024-dim image embedding for find-similar in the same pass (falling back to the older WD SwinV2 v3 if PixAI isn't installed); semantic search uses gte-multilingual-base to encode tags and your query into one semantic space, enabling cross-language searches like 'blue-toned girl' in Chinese. When tagging, English tags are replaced with their Chinese display names and written directly onto the Eagle element (English is kept only where no translation exists). Everything runs locally — no network, no billing, no Nepheline cost.

Cross-language search over your library. gte-multilingual-base (768-dim) encodes tags and your query into one semantic space, so a Chinese query like 'cyberpunk rainy night' still matches images tagged in English Danbooru tags. It searches tag text, then filters to tags that actually exist in your library.

The default image content-tagging model pixai-tagger-v0.9 (EVA02 architecture, ~1.2GB, character vocabulary through 2025-01), running locally on ONNX Runtime and harvesting a 1024-dim image embedding for find-similar in the same tagging pass. Falls back to the older WD SwinV2 v3 (SmilingWolf, ~467MB, vocabulary cutoff 2024-02) when PixAI isn't installed.

OpenAI's Contrastive Language-Image Pre-training model (ViT-B/32, 512-dim). Nephele uses it for zero-shot style classification and text-to-image search (find pictures in your local library that match an English description). Image-to-image similarity moved to EVA02 tagger embeddings in 2026-06 and no longer uses CLIP.

Microsoft's open-source ML inference engine. Nephele runs the PixAI Tagger, CLIP, the gte semantic model and others locally through it, no cloud required — preferring DirectML/CUDA acceleration and falling back to CPU without a GPU.

A topmost borderless window where you pin reference images and freely arrange, zoom, and rotate them. Lock mode makes it mouse-transparent so it doesn't block your canvas. Supports composition overlays, arrow annotations, and .pur file interop.

Double-press Ctrl or a mouse side button to summon a dual-ring radial menu arranged by artist workflow for instant access to any feature.

A full-screen eyedropper: press a hotkey (default Ctrl+Shift+I), click anywhere on screen, and the color is copied to the clipboard and written straight into 优动漫 / CLIP Studio Paint's foreground color. Windows-only.

Hosted-commissions extension (in development). Edit prices, open slots and a portfolio on the desktop, publish them as a link that's yours alone; requests from clients land straight in the desktop Commission Manager.

Mobile extension. Browse your Eagle library remotely, subscribe to artist updates, and chat with the assistant. Two modes: cloud-direct (works anywhere — Q&A / search / references) and MAX mode (requires your desktop online, remotely driving the desktop's full Agent (cloud tier) with local-library access and tools).

Chrome/Edge browser extension handling the last mile of publishing delivery and creator data-warehouse collection. Your login sessions stay in your own browser.

Aggregates your publish records, views, saves, and engagement across social platforms into a local data warehouse, paired with an Eagle-library portrait — so you no longer depend on each platform's scattered backend.

Scans image EXIF, IPTC, and generation-parameter fingerprints to detect signatures from Midjourney, ComfyUI, NovelAI, and more.

An open standard (Content Credentials) recording an image's provenance and edit history. AI Credential Check reads an image's C2PA manifest to tell whether it was AI-generated or processed by specific tools.

The internet timestamp protocol international standard. Nephele's digital certifications are backed by authorities like DigiCert.

The third-party authority that issues trusted timestamps. Nephele's certification sends only the file's 256-bit hash to a TSA (e.g. FreeTSA, DigiCert, IdenTrust) — your original never leaves your computer, and verification doesn't depend on Nephele's servers.

Cryptographic hash tree structure. In batch certification all files are chain-linked; tampering with any one breaks the entire tree.

Binary diff patch algorithm. Nephele's self-updater uses it to download only changed parts, drastically reducing update size.

UI automation & onboarding framework with real mouse/keyboard + QML tree observation + video recording. One script reused for testing, tutorials, and marketing.

Remote emergency shutdown. The server can issue commands to stop specific client versions from running, for security or compliance.

What Nephele calls its users. In Nephele's worldview, those who come to the Workshop to create are "Architects" — you lead the design and planning while the Agent does the heavy lifting, rather than laying every brick yourself.

A protection mode in Watermark & Credit: embed an invisible watermark into a work, then later extract it from a suspect image to verify whether it leaked from you and trace the repost chain. Usually survives light compression, resizing, and cropping.

When the full Agent tackles a complex task, it can split it via delegate_tasks into multiple Sub-Agents running in parallel (up to 6, each with a 120s timeout, cloud-tier only). A Sub-Agent can't delegate further (preventing infinite recursion), used to speed up parallelizable subtasks.

The default lens of Creative Review, focused on long-term growth rather than a single score: it gives concrete improvement suggestions and deliberately keeps you from fixating on an absolute number.

The numbers-driven lens in Creative Review: a gamified ten-tier letter grade, a dimension radar, and market price references. Benchmarked against the global spectrum of practitioners, so the bar is deliberately strict.

The full-screen preview you enter by double-clicking any image in the Resource Library: view the full image, zoom into details with the wheel anchored at the cursor, and step between images without leaving your browsing flow.

An identifier generated for device binding (the 2-device limit and theft detection): a UUID derived by one-way-hashing a few stable system fields — not your real hardware ID. Nephele doesn't collect your MAC / CPU ID / disk / BIOS serials; the server stores only the derived hash.

The AIGC label block that domestic AI services embed in files under China's national standard GB 45438-2025. AI Credential Check reads this structured field and maps the producer code to a readable name, treating it as a strong signal that the content is AI-generated.

Drop a PMX model file into Nephele and a standalone 3D preview window opens — mouse rotate/zoom/pan, a three-viewport mode, and lighting plus IBL ambient-light adjustment. Inspect a model's pose and structure without opening Blender or MMD, handy when drawing from 3D references.