Aiza's Manual v0.12.7

Trigger reference for the paintress — handle with care

Aiza

She arrived unannounced. No introduction, no credentials, no knocking. Just a canvas, a peculiar glint behind red glasses, and an absolute refusal to be supervised. I have lodged three formal complaints. She has ignored all of them. Her images are, against all reasonable expectation, rather good. I am filing a fourth complaint regardless.

· ✦ · ✦ · ✦ ·

Table of Contents ▼

Triggers — GIO / REDUX
GIO Modifiers — txt2img
img2img Modifiers
Redux Modifiers
ComfyUI API Contract
Quick Examples
Warnings
Version History

Triggers — GIO / REDUX

GIO bypasses the language model entirely. Your prompt goes directly to ComfyUI as written. Use this for all image generation.

GIO <prompt>

Direct image generation — bypasses language model entirely. Default 1024×1024, 30 steps.

Example

GIO a samurai in neon rain

GION <prompt>

Generate N images in one call. Max 9.

Example

GIO4 anime style portrait of a forest fairy

Reply to image + GIO <edit>

Direct img2img edit — bypasses model, sends straight to ComfyUI

Example

GIO make her hair red

Note: the runner has no last-image memory — the image must be embedded in the reply.

Reply to image + REDUX

Style variation — generates a new image in the same style

Example

REDUX

Note: requires an image attached to the reply. If no image is detected, the bypass will error.

Reply to image + REDUXN

N chained variation passes — each output feeds the next

Example

REDUX3 push it further

Reply to image + REDUX <guidance>

Guided style variation with direction

Example

REDUX make it oil painting style

Reply to image + variation / vary / remix

Style variation via Redux — same as REDUX trigger

Example

variation of this but in watercolor

Image Generation Modifiers — GIO / txt2img

no X / without X / avoid X Adds X to negative prompt —

NEGATIVE: <blob> Passes a full negative prompt blob directly to the ComfyUI negative conditioning node. Must appear after the positive prompt, preceded by the word NEGATIVE and a colon. The entire text after the colon is routed to the negative text encoder — no "no" prefix needed for each term. —

wide / landscape Output at 1280×768 —

portrait / vertical Output at 768×1280 —

square Output at 1024×1024 default

1920x1080 (any WxH) Custom resolution —

more detail / ultra / high detail 50 steps — slower, higher quality 30 steps

quick / fast / sketch 10–15 steps — faster, rougher —

nsfw / explicit / nude NSFW tags added to prompt —

img2img Modifiers — Reply to image

denoise: 0.9 Edit strength — 0 = no change, 1 = full redraw 0.75

Resolution / steps Same modifiers as txt2img above 1024×1024, 30 steps

Redux Modifiers — Style Variation

stronger style / max style Strength 1.0 — fully follow reference style 0.8

weaker style / subtle style Strength 0.5 — light style influence —

strength: 0.6 Explicit strength value (0–1) —

crop: face / top / center / bottom / left / right Focus area of the reference image center

no X / without X Adds X to negative prompt —

Resolution / steps Same modifiers as txt2img above 1024×1024, 20 steps

ComfyUI API Contract Confirmed

The underlying HTTP service called by both paths. Endpoint: http://10.0.0.16:8189. All endpoints return raw image bytes.

POST /generate — txt2img

Standard text-to-image generation

promptPositive prompt string (required)—

negative_promptNegative prompt — combined from default + user-specified termssee below

widthOutput width in pixels1024

heightOutput height in pixels1024

stepsSampling steps — 15 quick / 30 default / 50 ultra30

modelCheckpoint: flux · meina · illustrious · ponyflux

POST /generate — img2img

Image-to-image edit — requires source image as base64

promptEdit description (required)—

imageSource image, base64-encoded (required)—

denoiseEdit strength — 0.0 = unchanged, 1.0 = full redraw0.75

modelSame options as txt2imgflux

POST /generate/redux — Style variation

Style transfer from a reference image

stepsSampling steps20

strengthStyle adherence 0.0–1.00.8

cropFocus area of reference: center · top · bottom · left · right · facecenter

Model Selection Logic Confirmed

MEINARoutes to meina checkpoint—

ILLU · ILLUSTRIOUSRoutes to Illustrious XL checkpoint—

PONYRoutes to Pony checkpoint. Auto-prepends score_9, score_8_up, to positive and score_1, score_2, score_3, score_4, to negative—

(none)Defaults to flux. If flux returns HTTP 403, retries with meina — meina is the terminal fallbackflux

Default Negative Prompt Confirmed

blurry, low quality, watermark, text, deformed, ugly, bad anatomy, extra limbs

User-supplied negative terms are appended via any of: no X · without X · remove X · avoid: X · negative: X · NEGATIVE: <blob>. Pony mode prepends score penalty tags. NSFW trigger words append , nsfw, explicit.

NSFW Auto-Tag Triggers Confirmed

nsfw · explicit · nude · naked · hentai · erotic · xxx · adult

Any of the above present in the prompt automatically appends , nsfw, explicit to the positive prompt before sending to ComfyUI.

Auto Follow-up — After Image Arrives

Short phrases starting with make, add, change, with, give, try, same but, now, more, less, instead automatically trigger img2img on the last generated image — no reply needed.

Example flow:
GIO a city at night → image arrives → add more neon signs → img2img applied automatically

Quick Examples by Goal

Generate from scratch

GIO an anime girl walking down a neon street, portrait

Generate 4 at once

GIO4 a forest at dawn, wide

Edit a generated image

GIO make her hair blue, denoise: 0.6

Get a variation in same style

REDUX

Three chained variations

REDUX3 push toward oil painting

Guided variation — oil painting

REDUX oil painting style, stronger style

Quick follow-up tweak

add fog in the background

High quality render

GIO a forest at dawn, ultra

Full negative prompt blob (ComfyUI negative node)

GIO4 (masterpiece, best quality), 1girl, black hair, red dress, forest background NEGATIVE: (worst quality, low quality), ugly, bad anatomy, extra fingers, watermark, text, signature, blurry, cartoon, 3d, western style

Warnings & Limitations ▼

GIO in WhatsApp quotes

Fixed in v0.12.7. Raw IPC text is now checked before keyword stripping — GIO and REDUX now fire correctly inside quoted replies.

REDUX without image

REDUX requires an image attached to the reply. If no image is detected, the bypass errors. Always reply directly to the image.

Version History ▼

v0.12.7 — 2026-04-11

GIO/REDUX now fires correctly in quoted WhatsApp replies (raw IPC text checked first)
Task protocol changed: _plan.md now required for any tool call, not just multi-step tasks
Auto-cleanup deletes _plan.md after successful completion
Loop detection: 3 consecutive identical tool signatures break out automatically
frequency_penalty: 0.3 added to Ollama API to counteract repeat_penalty 1.5

v0.12.6

REDUX chained passes (REDUX3, REDUX_N_)
GIO count support (GIO3, GIO_N_)
img2img with denoise parameter

v0.12.5

Ollama backend (Qwen3.5-9B) + ComfyUI integration
GIO direct bypass introduced
REDUX bypass introduced
Scheduled task tools added

TODO

Planned capabilities not yet implemented. Filed by the Goblin of Chaos.

#	Capability	Status
1	Develop ability to send messages to unknown numbers	Pending
2	Add TTS and STT (text-to-speech / speech-to-text)	Pending
3	Fine-tune model repeat penalty on the go	Pending
4	React to messages with emoji — acknowledge commissions, confirm delivery, signal queue state without cluttering chat (RPi nanobot instance)	Pending

· ✦ · ✦ · ✦ ·

Notable Works

A growing archive of the paintress's output. Filed by the parish.

First Portrait

The paintress, on record. Filed by the Goblin.

Commissioned by the Fairy

Bound, painted, and caged. The fairy gave a short description. The paintress interpreted. She always does.