Ghibli Diffusion: the ghibli style token and how to use it

Q: Faces look distorted

512-pixel base and small faces in wide shots. Make the figure small and distant, or inpaint the face at a higher resolution after upscaling.

Q: Output looks like generic SD 1.5

Trigger phrase missing or buried late in the prompt. Put ghibli style first and raise CFG to 8.

Topic: Models

By Captain

Published 2026-09-30

Nitrosocke's Ghibli Diffusion was one of the first style models for Stable Diffusion. Learn its trigger phrase, settings, how to describe scenes it can draw, and creative ideas for painterly animation-style landscapes.

Overview

Ghibli

Diffusion by Nitrosocke is a Stable Diffusion 1.x fine-tune trained on stills in the style of the Japanese animation studio. It is triggered by the phrase ghibli style and excels at lush landscapes, cozy interiors, skies and gentle characters.

It knows the look, not the characters: there is no training on specific names, so describe appearances and settings. It is a 512-pixel model, so upscale with an anime-friendly

upscaler

for finals.

Reference

Name	Type	What it is
Trigger	ghibli style	Put it at the start of the prompt.
Prompt	text + tags	Short descriptive phrases work well; describe the scene, time of day and weather.
CFG	7	Standard SD 1.5 value.
Steps	25-30	Euler a or DPM++ 2M Karras.
Resolution	512x512 to 512x768	Upscale afterward; do not render large directly.

No install required - run AI models on the cloud

BitVector Prism is the easiest way to start: pick a model, type a prompt and generate in a clean web app, with nothing to configure. BitVector is also available on Discord and on the web (SpyGlass).

Try BitVector

Step by step

Begin with ghibli style, then the scene in one or two phrases.
Add time of day and weather words: golden hour, after the rain, summer haze.
Keep the
negative prompt
simple: lowres, blurry, text, watermark, bad anatomy.
Generate at 512x768 for vertical scenes or 768x512 for landscapes.
Upscale 2x with an anime upscaler and a light
img2img
pass at 0.35 denoise.

Examples

Landscape

ghibli style, a grassy hill with a lone tree, a girl in a yellow dress flying a kite, cumulus clouds, summer afternoon, distant sea

Interior

ghibli style, cozy bakery kitchen at dawn, bread in a wood oven, copper pots, flour dust in sunlight, an old baker smiling

Creative idea: sky harbor

ghibli style, a harbor town built on clouds, propeller boats docked at wooden piers, lanterns, sunset, children waving

Tips

Landscapes with one small figure are its signature; describe the figure briefly.
Flying machines, trains and seaside towns produce especially convincing results.
Interiors want warm light and clutter: kitchen, pots, hanging herbs, open window.
For characters, describe hair, clothing and expression; avoid celebrity or franchise names.

Troubleshooting

Faces look distorted

Why it happens

512-pixel base and small faces in wide shots.

How to fix it

Make the figure small and distant, or inpaint the face at a higher resolution after upscaling.

Output looks like generic SD 1.5

Why it happens

Trigger phrase missing or buried late in the prompt.

How to fix it

Put ghibli style first and raise

CFG

to 8.

No install required - run AI models on the cloud

PirateDiffusion is Telegram only and built for pros: thousands of models, LoRAs and workflows driven by chat commands, with unlimited generation on a fixed price plan.