Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,82 @@
|
|
| 1 |
---
|
| 2 |
license: creativeml-openrail-m
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: creativeml-openrail-m
|
| 3 |
+
language:
|
| 4 |
+
- en
|
| 5 |
+
library_name: diffusers
|
| 6 |
+
pipeline_tag: text-to-image
|
| 7 |
+
tags:
|
| 8 |
+
- art
|
| 9 |
+
- anime
|
| 10 |
+
- stable diffusion
|
| 11 |
+
- openvino
|
| 12 |
+
- onnx
|
| 13 |
---
|
| 14 |
+
|
| 15 |
+
|
| 16 |
+
SoteMix V2 is trained at 1024x1536.
|
| 17 |
+
Don't use this model at 512x resolutions, it will be blurry.
|
| 18 |
+
Use SoteMix V1 for 512x instead.
|
| 19 |
+
|
| 20 |
+
|
| 21 |
+
Positive Prompts:
|
| 22 |
+
```
|
| 23 |
+
(masterpiece, best quality, highres), 1girl,
|
| 24 |
+
```
|
| 25 |
+
|
| 26 |
+
Negative Prompts:
|
| 27 |
+
```
|
| 28 |
+
(worst quality, low quality, lowres), zombie, interlocked fingers,
|
| 29 |
+
```
|
| 30 |
+
|
| 31 |
+
Sampler: `Euler a`
|
| 32 |
+
Steps: `30-40`
|
| 33 |
+
Clip Skip: `1`
|
| 34 |
+
CFG: `7`
|
| 35 |
+
Base Resolution: `1024x` / `768x1280` / `1280x768` / `1024x1536` / `1920x1080`
|
| 36 |
+
Model can still be chaotic at `1024x1536` and `1920x1080`.
|
| 37 |
+
|
| 38 |
+
Second Pass / Hires:
|
| 39 |
+
Sampler: `Euler` / `Euler a`
|
| 40 |
+
Steps: `10` with `Euler` / `20` with `Euler a`
|
| 41 |
+
Upscaler: `RealESRGAN 4x+ Anime6B` / `ESRGAN 4x-AnimeSharp` with `0.3`-`0.4` denoise strength.
|
| 42 |
+
CFG: `11`
|
| 43 |
+
Resolution: `2x` of the base resolution.
|
| 44 |
+
|
| 45 |
+
|
| 46 |
+
|
| 47 |
+
Training:
|
| 48 |
+
My GPU couldn't handle full model training at these resolutions so i trained it as a `512` layer Lora with SoteMix V1 as the base.
|
| 49 |
+
Used highres as the trigger word. Also used raifu trigger word with my OC character.
|
| 50 |
+
|
| 51 |
+
Resolution: `1024x1536 with Bucketing`
|
| 52 |
+
Batch Size: `1`
|
| 53 |
+
Steps: `12000`
|
| 54 |
+
GPU: `Intel ARC A770 16GB`
|
| 55 |
+
|
| 56 |
+
Bucket:
|
| 57 |
+
```
|
| 58 |
+
bucket 0: resolution (832, 1664), count: 49
|
| 59 |
+
bucket 1: resolution (896, 1536), count: 1
|
| 60 |
+
bucket 2: resolution (960, 1536), count: 12
|
| 61 |
+
bucket 3: resolution (960, 1600), count: 492
|
| 62 |
+
bucket 4: resolution (1024, 1344), count: 112
|
| 63 |
+
bucket 5: resolution (1024, 1408), count: 31
|
| 64 |
+
bucket 6: resolution (1024, 1536), count: 768
|
| 65 |
+
bucket 7: resolution (1152, 1344), count: 6
|
| 66 |
+
bucket 8: resolution (1216, 1216), count: 25
|
| 67 |
+
bucket 9: resolution (1344, 1024), count: 16
|
| 68 |
+
bucket 10: resolution (1344, 1152), count: 2
|
| 69 |
+
bucket 11: resolution (1536, 896), count: 59
|
| 70 |
+
bucket 12: resolution (1536, 1024), count: 7
|
| 71 |
+
bucket 13: resolution (1600, 960), count: 82
|
| 72 |
+
bucket 14: resolution (1664, 832), count: 21
|
| 73 |
+
bucket 15: resolution (2304, 576), count: 1
|
| 74 |
+
```
|
| 75 |
+
|
| 76 |
+
Merge:
|
| 77 |
+
Merged SoteMix V1 with Lunar Radiance Light and then merged the Hires Lora i trained on top of it.
|
| 78 |
+
Merge ratio: `(0.6 SoteMix V1 + 0.4 Lunar Radiance Light) + 0.75 Hires Lora`
|
| 79 |
+
|
| 80 |
+
|
| 81 |
+
|
| 82 |
+
|