Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,73 @@
|
|
| 1 |
---
|
| 2 |
license: creativeml-openrail-m
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: creativeml-openrail-m
|
| 3 |
+
tags:
|
| 4 |
+
- text-to-image
|
| 5 |
+
- stable-diffusion
|
| 6 |
+
- anime
|
| 7 |
+
- aiart
|
| 8 |
---
|
| 9 |
+
|
| 10 |
+
**This model is trained on 6(+1?) characters from ONIMAI: I'm Now Your Sister! (γε
γ‘γγγ―γγγΎγ!)**
|
| 11 |
+
|
| 12 |
+
### Example Generations
|
| 13 |
+
|
| 14 |
+
|
| 15 |
+
|
| 16 |
+
|
| 17 |
+
### Usage
|
| 18 |
+
|
| 19 |
+
The model is shared in both diffuser and safetensors formats.
|
| 20 |
+
As for the trigger words, the six characters can be prompted with
|
| 21 |
+
`OyamaMahiro`, `OyamaMihari`, `HozukiKaede`, `HozukiMomiji`, `OkaAsahi`, and `MurosakiMiyo`.
|
| 22 |
+
`TenkawaNayuta` is tagged but she appears in fewer than 10 images so don't expect any good result.
|
| 23 |
+
There are also three different styles trained into the model: `aniscreen`, `edstyle`, and `megazine` (yes, typo).
|
| 24 |
+
As usual you can get multiple-character imagee but starting from 4 it is difficult.
|
| 25 |
+
|
| 26 |
+
In the following images are shown the generations of different checkpoints.
|
| 27 |
+
The default one is that of step 22828, but all the checkpoints starting from step 9969 can be found in the `checkpoints` directory.
|
| 28 |
+
They are all sufficiently good at the six characters but later ones are better at `megazine` and `edstyle` (at the risk of overfitting, I don't really know).
|
| 29 |
+
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
|
| 33 |
+
|
| 34 |
+
### More Generations
|
| 35 |
+
|
| 36 |
+
|
| 37 |
+
|
| 38 |
+
|
| 39 |
+
|
| 40 |
+
|
| 41 |
+
|
| 42 |
+
|
| 43 |
+
|
| 44 |
+
|
| 45 |
+
|
| 46 |
+
### Dataset Description
|
| 47 |
+
|
| 48 |
+
The dataset is prepared via the workflow detailed here: https://github.com/cyber-meow/anime_screenshot_pipeline
|
| 49 |
+
|
| 50 |
+
It contains 21412 images with the following composition
|
| 51 |
+
|
| 52 |
+
- 2133 onimai images separated in four types
|
| 53 |
+
- 1496 anime screenshots from the first six episodes (for style `aniscreen`)
|
| 54 |
+
- 70 screenshots of the ending of the anime (for style `edstyle`, not counted in the 1496 above)
|
| 55 |
+
- 528 fan arts (or probably some official arts)
|
| 56 |
+
- 39 scans of the covers of the mangas (for style `megazine`, don't ask me why I choose this name, it is bad but it turns out to work)
|
| 57 |
+
- 19279 regularization images which intend to be as various as possible while being in anime style (i.e. no photorealistic image is used)
|
| 58 |
+
|
| 59 |
+
Note that the model is trained with a specific weighting scheme to balance between different concepts so that every image does not weight equally.
|
| 60 |
+
After applying the per-image repeat we get around 145K images per epoch.
|
| 61 |
+
|
| 62 |
+
### Training
|
| 63 |
+
|
| 64 |
+
Training is done with [EveryDream2](https://github.com/victorchall/EveryDream2trainer) trainer with [ACertainty](https://huggingface.co/JosephusCheung/ACertainty) as base model.
|
| 65 |
+
The following configuration is used
|
| 66 |
+
|
| 67 |
+
- resolution 512
|
| 68 |
+
- cosine learning rate scheduler, lr 2.5e-6
|
| 69 |
+
- batch size 8
|
| 70 |
+
- conditional dropout 0.08
|
| 71 |
+
- change beta scheduler from `scaler_linear` to `linear` in `config.json` of the scheduler of the model
|
| 72 |
+
|
| 73 |
+
I trained for two epochs wheareas the default release model was trained for 22828 steps as mentioned above.
|