Update README.md
Browse files
README.md
CHANGED
|
@@ -1,10 +1,71 @@
|
|
| 1 |
---
|
| 2 |
-
license:
|
| 3 |
language:
|
| 4 |
- en
|
| 5 |
library_name: diffusers
|
| 6 |
pipeline_tag: text-to-image
|
| 7 |
tags:
|
| 8 |
- text-to-image
|
|
|
|
|
|
|
| 9 |
---
|
| 10 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
+
license: openrail++
|
| 3 |
language:
|
| 4 |
- en
|
| 5 |
library_name: diffusers
|
| 6 |
pipeline_tag: text-to-image
|
| 7 |
tags:
|
| 8 |
- text-to-image
|
| 9 |
+
base_model:
|
| 10 |
+
- stabilityai/stable-diffusion-xl-base-1.0
|
| 11 |
---
|
| 12 |
+
|
| 13 |
+
# Anime Stable Diffusion Model
|
| 14 |
+
|
| 15 |
+
A custom Stable Diffusion model fine-tuned for anime-style image generation, trained on a large dataset of anime images.
|
| 16 |
+
This is the first concept model for the entire series as I am spending more time filtering and processing the
|
| 17 |
+
larger dataset. Currently the model is still undertrained, while it can reflect certain notions, a lot of additional
|
| 18 |
+
improvements need to be done.
|
| 19 |
+
|
| 20 |
+
## Prompt
|
| 21 |
+
Danbooru style tagging.
|
| 22 |
+
|
| 23 |
+
Quality tag: Masterpiece, high quality, normal quality, low quality
|
| 24 |
+
Aesthetic tag: Very aesthetic, aesthetic, pleasent, unpleasent
|
| 25 |
+
|
| 26 |
+
Additional special tag: High resolution, elegant, artist:
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
| Rating Modifier | Rating Criterion |
|
| 30 |
+
| --------------- | ---------------- |
|
| 31 |
+
| - | general |
|
| 32 |
+
| - | sensitive |
|
| 33 |
+
| nsfw | questionable |
|
| 34 |
+
| nsfw | explicit |
|
| 35 |
+
|
| 36 |
+
Recommanded prompt order: Rating tag, quality tag, aesthetic tag, (additional tag), general tag
|
| 37 |
+
|
| 38 |
+
### Dataset Specifications
|
| 39 |
+
- Total Images: 172k
|
| 40 |
+
- General Training Set: 160k images
|
| 41 |
+
- Aesthetic Fine-tuning Set: 12k high-quality images
|
| 42 |
+
- Resolution: 1024x1024
|
| 43 |
+
|
| 44 |
+
### Hardware Configuration
|
| 45 |
+
- GPUs: 2x NVIDIA RTX 6000 Ada
|
| 46 |
+
- Training Time: 16 days (General), 3 days (Aesthetic fine tune)
|
| 47 |
+
|
| 48 |
+
### Training Configuration
|
| 49 |
+
|
| 50 |
+
| Parameter | Value | Description |
|
| 51 |
+
|-----------|--------|-------------|
|
| 52 |
+
| Resolution | 1024x1024 | Training resolution |
|
| 53 |
+
| Batch Size | 8x2x2 | Effective batch size |
|
| 54 |
+
| Learning Rate | 5e-5 | Base learning rate |
|
| 55 |
+
| Text Encoder LR | 1e-5 | Learning rate for text encoder |
|
| 56 |
+
| Epochs | 10 | Total training epochs |
|
| 57 |
+
| Mixed Precision | FP16 | Training precision mode |
|
| 58 |
+
| Optimizer | AdamW8bit | Optimizer type |
|
| 59 |
+
|
| 60 |
+
### Advanced Settings
|
| 61 |
+
|
| 62 |
+
| Feature | Setting | Purpose |
|
| 63 |
+
|---------|---------|----------|
|
| 64 |
+
| Gradient Checkpointing | Enabled | Memory optimization |
|
| 65 |
+
| XFormers | Enabled | Attention optimization |
|
| 66 |
+
| Memory Efficient Attention | Enabled | Memory optimization |
|
| 67 |
+
| Bucket Resolution Steps | 128 | Dynamic resolution handling |
|
| 68 |
+
| Min Bucket Resolution | 512 | Minimum image size |
|
| 69 |
+
| Max Bucket Resolution | 4096 | Maximum image size |
|
| 70 |
+
| Noise Offset | 0.035 | Training stability |
|
| 71 |
+
| Min SNR Gamma | 5 | Signal-to-noise ratio control |
|