File size: 2,351 Bytes
79f3200 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 | ---
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.
|