Presets
White Lotus Preset
SillyTavern Preset

White Lotus

A preset built to get LLMs out of their own way — more initiative, more creativity, fewer stalled scenes.

Tested on GLM, Kimi, Deepseek, and Gemma. Best suited for character-driven and plot-focused roleplay.

WHITE LOTUS [3.0.0].json
Latest  ·  SillyTavern Preset  ·  JSON
Download
Previous: v2.0.0 · v1.2.1 · v1.1.0
All hosted via NanoGPT unless noted. Post-processing: Merge consecutive roles (Semi-strict for GLM 5/5.1). Context: 35k set, ~20k practical with summary extension.
Kimi 2.5 Thinking Default model. My love.
Temp 1.0 · Freq 0 · Pres 0 · Top P 0.90 · Top K 0 · Rep P 1 · Min P 0.01 · Top A 0
Alt Temp 1.0 · Freq 0 · Pres 0 · Top P 0.95 · Top K 25 · Rep P 1 · Min P 0 · Top A 1
Kimi 2.5 (Non-Think)
Temp 0.6 · Freq 0 · Pres 0 · Top P 0.95 · Top K 0 · Rep P 1 · Min P 0.01 · Top A 0
Kimi 2.6
Temp 1.0 · Freq 0 · Pres 0 · Top P 1.0 · Top K 0 · Rep P 1 · Min P 0 · Top A 0
GLM 5 / 5.1 Semi-strict post-processing
Temp 1.0 · Freq 0–0.2 · Pres 0–0.2 · Top P 0.95 · Top K 40 · Rep P 1 · Min P 0 · Top A 0
Alt Temp 1.0 · Freq 0.02 · Pres 0.03 · Top P 1 · Top K 0 · Rep P 1 · Min P 0.05 · Top A 0.4
GLM 4.7
Temp 1.0 · Freq 0–0.2 · Pres 0–0.2 · Top P 0.95 · Top K 40 · Rep P 1 · Min P 0 · Top A 0
Deepseek 3.2 Exp Thinking — DS docs say most params are ignored
Temp 1.0 · Freq 0 · Pres 0 · Top P 0.95 · Top K 0 · Rep P 1 · Min P 0 · Top A 0
Deepseek 4 Flash Super sensitive to temp!
Temp 0.7 · Freq 0 · Pres 0 · Top P 0 · Top K 0 · Rep P 1 · Min P 0 · Top A 0
DS 4 Flash (Think)
DS 4 Pro
Temp 1.0 · Freq 0 · Pres 0 · Top P 1.0 · Top K 0 · Rep P 1 · Min P 0 · Top A 0
Gemma 4 31B
Temp 1.0 · Freq 0–0.2 · Pres 0–0.2 · Top P 0.95 · Top K 64 · Rep P 1 · Min P 0 · Top A 0

LLMs love to write beautiful prose. They'll set up vivid scenes, offer gorgeous internal commentary, describe the way light falls on a character's face like they're angling for a literary prize. But actually doing something? Introducing an unexpected element? Pushing the story forward without holding your hand? That's where they choke.

White Lotus is my attempt to fix that. The focus is on writing quality, adaptability, and keeping the token count reasonable — no bloated instruction walls. It's deliberately less "instruction-voiced" than most presets, which is a conscious tradeoff. The output sounds more human and less like a model following orders, even if it's slightly less token-efficient.

This preset is built for roleplay that cares about plot and characters. It's probably terrible for RPG-mechanic-style stuff — fair warning. But if you want NPCs that feel like actual people, a world that keeps moving whether or not you're paying attention, and characters who make choices based on what they want rather than what's convenient? That's what this is for.

See how White Lotus 3.0.0 handles the same scene across the Kimi, GLM, and Deepseek families — base runs vs. the Variety Toggle, each on its recommended parameters.
3.0.0 Tests
Older: v1.2.0 tests (image-based)
Core

The Lotus Engine handles the fundamentals: characters have lives outside of the current scene, make choices based on their own wants and needs, can lie or deceive when it makes sense, and develop slowly (with room to regress). It enforces knowledge boundaries — characters only know what they've seen or heard — and keeps the user's agency sacred.

Setup

Variables you set once and forget about: tense, POV, diction, length, and narrator type. Narrator can be Omniscient (knows everyone's inner world except the user's) or Character (becomes the character — biases, blind spots, and all). Diction options for Chinese, Japanese, or Historical settings add culturally appropriate idioms, honorifics, and metaphors.

Genre presets: Contemporary, Slice of Life, Southern Gothic, Science Fiction, Space Opera, Dark Fantasy, Regency Historical, LitRPG, Historical Chinese, Modern Chinese — each one sets the primary genre, subgenre, tone, and reference style in one toggle. There's also a companion extension that makes managing toggles and settings easier, with some bonus UI tweaks and chat styling.

Modules

Toggle on what you need, leave off what you don't. These are designed to be genuinely modular — no dependencies, no conflicts. Only turn on NSFW when things are getting frisky; only use Anti-Slop if your model needs it (looking at you, GLM).

Anti-Slop
Better NPCs
Disabilities
Consent
Negativity Bias
NSFW (Realistic)
NSFW (Gooner)
Intense Violence
Abuse / Toxic
Direction

Variety Toggle randomizes an additional story element each turn — a memory from the character's past, an intrusive thought, a sudden commotion, a misunderstanding being planted for later. Great for breaking repetitive loops. CoT (Brainstorm) prompts thinking models to work through character motivation, scene setup, and interpersonal dynamics before writing. There's also a Hyper-Mode for CoT that adds thematic reasoning on top.

Tools

NPC Crafter generates randomized character descriptions with specific physical traits, fashion sense, speaking style, personality, and a quirk — produces way more interesting characters than letting the model default to "tall, handsome, mysterious." Relationship Tracker and Temporal Tracker keep stats on hostility, interest, obligation, attraction, time, weather, and location. Status Board tracks each character's health, hunger, energy, hygiene, arousal, mood, and a private thought. If you use any trackers, toggle on Tracker Rules to keep the output format consistent. All come with styled regex scripts for the UI.