Update README.md
Browse files
README.md
CHANGED
|
@@ -8,23 +8,26 @@ license: creativeml-openrail-m
|
|
| 8 |
|
| 9 |
This is a low-quality bocchi-the-rock (ぼっち・ざ・ろっく!) character model.
|
| 10 |
Similar to my [yama-no-susume model](https://huggingface.co/alea31415/yama-no-susume), this model is capable of generating **multi-character scenes** beyond images of a single character.
|
| 11 |
-
Of course, the result is still hit-or-miss, but I
|
| 12 |
and otherwise, you can always rely on inpainting.
|
| 13 |
Here are two examples:
|
| 14 |
|
| 15 |
With inpainting
|
| 16 |
-
|
| 17 |
|
| 18 |
Without inpainting
|
| 19 |
-
|
| 20 |
|
| 21 |
|
| 22 |
### Characters
|
| 23 |
|
| 24 |
The model knows 12 characters from bocchi the rock.
|
| 25 |
-
The ressemblance with a character can be improved by a better description of their appearance.
|
|
|
|
|
|
|
|
|
|
|
|
|
| 26 |
|
| 27 |
-
*Coming soon*
|
| 28 |
|
| 29 |
### Dataset description
|
| 30 |
|
|
@@ -51,7 +54,7 @@ The model is trained on runpod using 3090 and cost me around 15 dollors.
|
|
| 51 |
|
| 52 |
#### Hyperparameter specification
|
| 53 |
|
| 54 |
-
|
| 55 |
|
| 56 |
Note that as a consequence of the weighting scheme which translates into a number of different multiply for each image,
|
| 57 |
the count of repeat and epoch has a quite different meaning here.
|
|
@@ -61,16 +64,33 @@ and therefore I did not even finish an entire epoch with the 48000 steps at batc
|
|
| 61 |
### Failures
|
| 62 |
|
| 63 |
- For the first 24000 steps I use the trigger words `Bfan1` and `Bfan2` for the two fans of Bocchi.
|
| 64 |
-
However, these two words are too similar and the model fails to different characters for these.
|
|
|
|
|
|
|
|
|
|
|
|
|
| 65 |
|
| 66 |
|
| 67 |
### More Example Generations
|
| 68 |
|
| 69 |
With inpainting
|
| 70 |
-
|
|
|
|
|
|
|
|
|
|
| 71 |
|
| 72 |
Without inpainting
|
| 73 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 74 |
|
| 75 |
Some failure cases
|
| 76 |
-
|
|
|
|
|
|
|
|
|
|
|
|
| 8 |
|
| 9 |
This is a low-quality bocchi-the-rock (ぼっち・ざ・ろっく!) character model.
|
| 10 |
Similar to my [yama-no-susume model](https://huggingface.co/alea31415/yama-no-susume), this model is capable of generating **multi-character scenes** beyond images of a single character.
|
| 11 |
+
Of course, the result is still hit-or-miss, but I with some chance you can get the entire Kessoku Band right in one shot,
|
| 12 |
and otherwise, you can always rely on inpainting.
|
| 13 |
Here are two examples:
|
| 14 |
|
| 15 |
With inpainting
|
| 16 |
+
|
| 17 |
|
| 18 |
Without inpainting
|
| 19 |
+
|
| 20 |
|
| 21 |
|
| 22 |
### Characters
|
| 23 |
|
| 24 |
The model knows 12 characters from bocchi the rock.
|
| 25 |
+
The ressemblance with a character can be improved by a better description of their appearance (for example by adding long wavy hair to ShimizuEliza).
|
| 26 |
+
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
|
| 30 |
|
|
|
|
| 31 |
|
| 32 |
### Dataset description
|
| 33 |
|
|
|
|
| 54 |
|
| 55 |
#### Hyperparameter specification
|
| 56 |
|
| 57 |
+
The model is trained for 48000 steps, at batch size 4, lr 1e-6, resolution 512, and conditional dropping rate of 10%.
|
| 58 |
|
| 59 |
Note that as a consequence of the weighting scheme which translates into a number of different multiply for each image,
|
| 60 |
the count of repeat and epoch has a quite different meaning here.
|
|
|
|
| 64 |
### Failures
|
| 65 |
|
| 66 |
- For the first 24000 steps I use the trigger words `Bfan1` and `Bfan2` for the two fans of Bocchi.
|
| 67 |
+
However, these two words are too similar and the model fails to different characters for these.
|
| 68 |
+
Therefore I changed Bfan2 to Bofa2 at step 24000. This seemed to solve the problem.
|
| 69 |
+
- Character blending is always an issue.
|
| 70 |
+
- When prompting the four characters of Kessoku Band we often get side shots.
|
| 71 |
+
I think this is because of some overfitting to a particular image.
|
| 72 |
|
| 73 |
|
| 74 |
### More Example Generations
|
| 75 |
|
| 76 |
With inpainting
|
| 77 |
+
|
| 78 |
+
|
| 79 |
+
|
| 80 |
+
|
| 81 |
|
| 82 |
Without inpainting
|
| 83 |
+
|
| 84 |
+
|
| 85 |
+
|
| 86 |
+
|
| 87 |
+
|
| 88 |
+
|
| 89 |
+
|
| 90 |
+
|
| 91 |
|
| 92 |
Some failure cases
|
| 93 |
+
|
| 94 |
+
|
| 95 |
+
|
| 96 |
+
|