---
license: openrail++
base_model: OnomaAIResearch/Illustrious-XL-v1.1
pipeline_tag: text-to-image
tags:
  - stable-diffusion-xl
  - text-to-image
  - diffusers
  - anime
library_name: diffusers
---

# Illustrious XL v1.1 — diffusers layout

A format conversion of [**OnomaAIResearch/Illustrious-XL-v1.1**](https://huggingface.co/OnomaAIResearch/Illustrious-XL-v1.1),
by OnomaAI Research. **All credit for the model belongs to them.**

The original is published as a single CompVis-layout `.safetensors`. This repository holds the
same weights rearranged into the standard diffusers directory layout, so it can be loaded without
a conversion step. Nothing was retrained, pruned, or quantized.

## Licence

**CreativeML Open RAIL++-M**, carried over unchanged from the original release. Commercial use is
permitted, subject to the use restrictions in the licence. Read it on the
[original model page](https://huggingface.co/OnomaAIResearch/Illustrious-XL-v1.1).

## What was done

Converted with `SDXLSingleFileConverter` from
[mindfire-image](https://github.com/Gatcha-man/mindfire-image): 2641 tensors written
(unet 1680, text_encoder_2 517, vae 248, text_encoder 196). Every tensor's shape was verified
against Stability's SDXL 1.0 base, which also supplied the configs and tokenizers a single file
does not carry. One key was dropped, `conditioner.embedders.1.model.logit_scale`, which diffusers'
text encoder does not use.

## Usage

OnomaAI's recommended settings: **Euler a, 20–28 steps, CFG 5–7.5**, at SDXL's native 1024²
(booru framings like 832×1216 cost no more).

Illustrious is trained on **Danbooru tags**, not sentences, and learned an explicit quality
vocabulary — `worst quality → bad → average → good → best quality`, plus `masterpiece`.

```
1girl, solo, long hair, white dress, looking at viewer, outdoors, masterpiece, best quality
```

Suggested negative prompt, from the model card:

```
worst quality, comic, multiple views, bad quality, low quality, lowres, displeasing,
very displeasing, bad anatomy, bad hands, scan artifacts, monochrome, greyscale
```

## Note on LoRAs

Most Illustrious LoRAs on Civitai target **v0.1**, not v1.1. Same architecture and tag vocabulary,
different weights — they generally still fire, but expect to lower the scale and to see some drift
from the samples they advertise.