NovelAI

Disclaimer

This specific page uses temporary drafts written and maintained by Generative AI. While verified to be accurate, please cross-verify with source code.

NovelAI connects text models (GLM 4.6, Xialong, Kayra, Erato) and anime image generation. You can use NovelAI for both, or pair its image generation with another chat provider.

Text models and subscription

GLM 4.6 is the default text model. Xialong and Erato require an active Opus subscription. Selecting Xialong checks your saved server or personal key; if you select it before adding a key, the check runs when you add the key. The serving key’s subscription determines the context allowance. Server managers can check Opus generation allowance with /novelai usage.

NovelAI’s Chat Format guide describes the story-style approach used for Kayra and Erato; GLM 4.6 and Xialong use separate chat templates. The filtering and polishing below are TomoriBot’s adaptations.

Sampling presets

Switching text models loads Carefree-Kayra, Erato-Shosetsu, or Default-Xialong for the selected model; GLM 4.6 uses its own defaults. Switching resets temperature, top-k, top-p, and min-p and enables those samplers for both server and personal model switches. Change presets or samplers in /config > Models > Text Samplers & Parameters; later edits stay active, and replacing a key on the same model preserves them.

Context and settings

Omitted settings stay saved in /config > Behavior > General Behavior for other providers.

Content or setting	NovelAI handling
Persona attributes, sample dialogue, conversation, and memories	Kept, including visible personal and short-term memories. Privacy and memory-tag filters still apply.
Lorebooks	Kept as entry contents. Always On lore stays with persistent information; activated lore sits above the last four speaker turns.
System, persona, and channel prompts; context notes; humanizer rules; conditioning; SillyTavern instructions	Omitted. Xialong uses NovelAI’s system prompt.
Status, time/channel/server info, documents, and emoji/sticker/sprite inventories	Omitted.
Memory-tool nudges	Omitted, including nudges built before a personal override or fallback.
ATTG metadata and quality stars	ATTG means Author, Title, Tags, and Genre. Xialong puts metadata first in its persistent information and defaults to 4 stars. Erato uses configured stars; Kayra omits stars.
Response prefills	Kept when allowed by the prefill controls.
Tools and media	No text-model tool calls or image/video input. Slash commands and separate image-generation commands remain available.
Token limits	Counted locally with room reserved for replies. Configured reply length takes priority over preset length and may be reduced to fit.

Known foreign speaker labels and standalone --- or *** lines stop output, even with the shared speaker-stop setting off. The responding persona’s names, aliases, and sprite labels remain allowed. The local guard preserves code; literal API stops can still match inside generated code.

Context trimming

Memories and lore share up to 20% of the context window. Older memories leave first; undated entries use their source order. Always On lore can also be trimmed. If the prompt still does not fit, TomoriBot trims it in this order:

Order	Content removed	Content protected
1	Extra sample dialogue	One available sample exchange
2	Memories and lore	No minimum
3	Old conversation history	Up to two recent history exchanges

Persona attributes, ATTG metadata, and the current turn also stay protected. Generation fails if the protected context cannot fit. Trimming leaves saved memories, lore, and messages unchanged.

/tool prompt snapshot shows the final prompt, token counts, reply cap, and trimming report.

Text polishing and Discord delivery

Polishing changes only the copy of history, memories, lore, and sample dialogue sent to all four NovelAI models, before token counting. Saved text and other providers are unchanged; grammar is not rewritten.

• Text is properly capitalized: Latin sentence and line starts are capitalized; known persona names use their configured casing.
• Prose lines receive punctuation: Missing final periods are added while existing punctuation, line breaks, and blank lines stay in place.
• Formatted prose is polished: *kisses juno* becomes *Kisses Juno.*. Bold, underline, strikeout, and spoilers follow the same rules; inline emphasis keeps its casing except for names.
• Code and markup are preserved: Code, URLs, mentions, Discord markup, lists, headings, blockquotes, tables, and separators keep their contents or structure.
• Non-prose starts and endings are respected: Lines starting with code fences, URLs, mentions, emoji, or non-Latin letters skip sentence formatting. Emoji, kaomoji, code, and link endings gain no extra period.
• Long-term memories use plain contents: Added owner labels, database IDs, and tag prefixes are omitted; resolved names remain.
• Generated history annotations are removed: Reply references, reaction metadata, and administrative notices are omitted. Gestures become actions; scenes and summaries lose their system wrapper. Literal user text containing System: is retained.
• Prefills stay exact: Explicit response prefills and continuation fragments are unchanged.
• Discord delivery follows your settings: The humanizer still controls casing, punctuation, and message splitting. Later history is polished again; lost punctuation cannot be recovered exactly.

Image generation

Assign a model to the NovelAI Image slot under /config > Models > Switch Models. Members can generate images directly with /novelai generate image, or ask in chat when the active text provider supports tool calling.

The Default + Image Generation Expert preset uses imageboard-style tags for NovelAI, with short descriptions when needed for interactions or text placement. For scenes with multiple characters, TomoriBot keeps shared scenery and camera directions separate from each character’s appearance and actions.

For readable speech, ask TomoriBot to include the exact dialogue. The NovelAI image tool places character dialogue in a trailing Text: section automatically. V4 and V4.5 support English text; V5 also supports Japanese and Chinese. Text on signs or screens can be described in the shared scene prompt. Keep requested text short and avoid no text among positive tags. Placement and typography (including vertical Japanese writing) are suggestions rather than guarantees.