Update README.md
Browse files
README.md
CHANGED
|
@@ -1,5 +1,66 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: other
|
| 3 |
-
license_name: fair-ai-public-license-1.0-sd
|
| 4 |
-
license_link: https://freedevproject.org/faipl-1.0-sd/
|
| 5 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: other
|
| 3 |
+
license_name: fair-ai-public-license-1.0-sd
|
| 4 |
+
license_link: https://freedevproject.org/faipl-1.0-sd/
|
| 5 |
+
datasets:
|
| 6 |
+
- KBlueLeaf/danbooru2023-webp-4Mpixel
|
| 7 |
+
- KBlueLeaf/danbooru2023-sqlite
|
| 8 |
+
language:
|
| 9 |
+
- en
|
| 10 |
+
library_name: diffusers
|
| 11 |
+
pipeline_tag: text-to-image
|
| 12 |
+
---
|
| 13 |
+
|
| 14 |
+
# Kohaku XL Epsilon rev2
|
| 15 |
+
join us: https://discord.gg/tPBsKDyRR5
|
| 16 |
+
|
| 17 |
+
## Rev2 Features
|
| 18 |
+
- Resumed from Kohaku XL Epsilon rev1
|
| 19 |
+
- 1.56M images, 5epoch
|
| 20 |
+
- Trained on selected artists' artworks and images about selected series/games
|
| 21 |
+
- Trained on PVC figure photos, can generate PVC style without any additional models
|
| 22 |
+
|
| 23 |
+
## Usage (PLEASE READ THIS SECTION)
|
| 24 |
+
### Prompt Format
|
| 25 |
+
`<1girl/1boy/1other/...>, <character>, <series>, <artists>, <general tags>, <quality tags>, <year tags>, <meta tags>, <rating tags>`
|
| 26 |
+
|
| 27 |
+
### Special Tags
|
| 28 |
+
- Quality tags: masterpiece, best quality, great quality, good quality, normal quality, low quality, worst quality
|
| 29 |
+
- Rating tags: safe, sensitive, nsfw, explicit
|
| 30 |
+
- Date tags: newest, recent, mid, early, old
|
| 31 |
+
|
| 32 |
+
#### Rating tags
|
| 33 |
+
General: safe
|
| 34 |
+
Sensitive: sensitive
|
| 35 |
+
Questionable: nsfw
|
| 36 |
+
Explicit: nsfw, explicit
|
| 37 |
+
|
| 38 |
+
### Resolution
|
| 39 |
+
This model is trained for resolutions from ARB 1024x1024 with minimum resolution 256 and maximum resolution 4096. This means you can use the standard SDXL resolution. However, opting for a slightly higher resolution than 1024x1024 is recommended. Applying a hires-fix is also suggested for better results.
|
| 40 |
+
|
| 41 |
+
## Training
|
| 42 |
+
- Hardware: Quad RTX 3090s
|
| 43 |
+
- Num Train Images: 1,536,902
|
| 44 |
+
- Total Epoch: 5
|
| 45 |
+
- Total Steps: 15015
|
| 46 |
+
- Training Time: 410 hours (wall time)
|
| 47 |
+
- Batch Size: 4
|
| 48 |
+
- Grad Accumulation Step: 32
|
| 49 |
+
- Equivalent Batch Size: 512
|
| 50 |
+
- Optimizer: Lion8bit
|
| 51 |
+
- Learning Rate: 1e-5 for UNet / 2e-6 for TE
|
| 52 |
+
- LR Scheduler: Cosine (with warmup)
|
| 53 |
+
- Warmup Steps: 1000
|
| 54 |
+
- Weight Decay: 0.1
|
| 55 |
+
- Betas: 0.9, 0.95
|
| 56 |
+
- Min SNR Gamma: 5
|
| 57 |
+
- Noise Offset: 0.0357
|
| 58 |
+
- Resolution: 1024x1024
|
| 59 |
+
- Min Bucket Resolution: 256
|
| 60 |
+
- Max Bucket Resolution: 4096
|
| 61 |
+
- Mixed Precision: FP16
|
| 62 |
+
- Caption Tag Dropout: 0.2
|
| 63 |
+
- Caption Dropout: 0.05
|
| 64 |
+
|
| 65 |
+
## License:
|
| 66 |
+
Fair-AI-public-1.0-sd
|