Update README.md
Browse files
README.md
CHANGED
|
@@ -2,6 +2,22 @@
|
|
| 2 |
license: apache-2.0
|
| 3 |
---
|
| 4 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 5 |
## Training
|
| 6 |
|
| 7 |
```python
|
|
|
|
| 2 |
license: apache-2.0
|
| 3 |
---
|
| 4 |
|
| 5 |
+
## DiT-B
|
| 6 |
+
|
| 7 |
+
Read the [blog](https://ruwwww.github.io/al-folio/blog/2026/waifu-diffusion/) about the base model.
|
| 8 |
+
|
| 9 |
+
## What's new
|
| 10 |
+
|
| 11 |
+
- Flux2 RoPE calculation for both the image and text sequences
|
| 12 |
+
- [Text embeddings](https://huggingface.co/nebulette/booru-character-aware-tokenizer) as a condition
|
| 13 |
+
- [Klein-like](https://huggingface.co/nebulette/klein-x), bias-free modules and QK norm; Z-Image-like, shift-free AdaLN
|
| 14 |
+
- Initialized from the [base model](https://huggingface.co/ruwwww/waifu_diffusion), kept the CIELAB output
|
| 15 |
+
|
| 16 |
+
## Future
|
| 17 |
+
|
| 18 |
+
- Sequence separated spatial separation built into the tokenizer
|
| 19 |
+
- Masked image training with the spatial tokenizer, leave the concept bleeding behind
|
| 20 |
+
|
| 21 |
## Training
|
| 22 |
|
| 23 |
```python
|