Update README.md
Browse files
README.md
CHANGED
|
@@ -4,31 +4,34 @@
|
|
| 4 |
|
| 5 |
|
| 6 |
----
|
| 7 |
-
# Diffusion Design Works
|
| 8 |
<span style="font-size: 70%;">This page is for historical and developmental progression purposes.
|
| 9 |
</span>
|
| 10 |
-
Welcome to the repository page of Diffusion Design Works!
|
| 11 |
|
|
|
|
| 12 |
Diffusion Design Works is a Stable Diffusion model series created with the goal of producing AI Art Generations with illustration styles mimicking the look and quality of studio produced Japanese animations.
|
| 13 |
|
| 14 |
What separates this model project from other checkpoints that similarly go about recreating popular artists’ and media’s styles lies mostly on a larger, higher fidelity dataset and an emphasis on doing a full finetune training rather than training a LoRA based network that then gets merged into an SD1.x, Novel AI, or SDXL base checkpoint, giving it better flexibility with the use of other extensions such as Controlnet or AnimateDiff that may struggle to incorporate generating with additional networks enabled.
|
|
|
|
| 15 |
----
|
| 16 |
|
|
|
|
|
|
|
| 17 |
Unlimited Diffusion Works is a model based on the works of animation studio ufotable, whos works including 14 Animated Films, 15 TV Shows, and Opening Animations and Cutscenes for 17 Video Game works for several different franchises, giving a diverse palate of different animators, styles, settings, and subjects while having a recognizable visual finish due to their CGI usage and consistent color grading technique.
|
| 18 |
|
| 19 |
This model’s current version was trained on 09/2023, with around 200k images over NovelAI v1, uses the following content from studio ufotable:
|
| 20 |
-
Fate Zero (TV)
|
| 21 |
-
The Garden of Sinners 8 – Epilogue (Movie)
|
| 22 |
-
The Garden of Sinners 9 - Future Gospel (Movie)
|
| 23 |
-
The Garden of Sinners 9.5 - Future Gospel: Extra Chorus (Movie – 31m 53s)
|
| 24 |
-
Fate Stay Night [Réalta Nua] (VG – 4m 30s)
|
| 25 |
-
Fate Stay Night Unlimited Blade Works (TV)
|
| 26 |
-
Fate/Hollow Ataraxia (VG – 3m)
|
| 27 |
-
Fate Stay Night Heaven’s Feel Trilogy (Movies – 5h 59m 13s)
|
| 28 |
|
| 29 |
In addition to ufotable, I added for experimentation purposes the dataset from the following non ufotable works for additional data and regularization purposes:
|
| 30 |
-
Lord El-Melloi II Case Files (TV – Studio Troyca)
|
| 31 |
-
*and additional miscellaneous non-animation frame data not from danbooru*
|
| 32 |
|
| 33 |
Dataset was created by running Blu-ray rips of the listed content through ffmpeg as part of the aid of Anime-Screenshot-Pipeline on Github back in February of 2023, but I had quickly become dissatisfied with the quality of the dataset it originally produced, and parts of its pipeline were not useful for what I needed, thus I took some tools, repurposed and/or modified settings, and redid the entire dataset after several months of doing initial training tests before committing to an overall.
|
| 34 |
|
|
|
|
| 4 |
|
| 5 |
|
| 6 |
----
|
| 7 |
+
## Diffusion Design Works
|
| 8 |
<span style="font-size: 70%;">This page is for historical and developmental progression purposes.
|
| 9 |
</span>
|
|
|
|
| 10 |
|
| 11 |
+
Welcome to the repository page of Diffusion Design Works!
|
| 12 |
Diffusion Design Works is a Stable Diffusion model series created with the goal of producing AI Art Generations with illustration styles mimicking the look and quality of studio produced Japanese animations.
|
| 13 |
|
| 14 |
What separates this model project from other checkpoints that similarly go about recreating popular artists’ and media’s styles lies mostly on a larger, higher fidelity dataset and an emphasis on doing a full finetune training rather than training a LoRA based network that then gets merged into an SD1.x, Novel AI, or SDXL base checkpoint, giving it better flexibility with the use of other extensions such as Controlnet or AnimateDiff that may struggle to incorporate generating with additional networks enabled.
|
| 15 |
+
|
| 16 |
----
|
| 17 |
|
| 18 |
+
# Unlimited Diffusion Works (UDW)
|
| 19 |
+
|
| 20 |
Unlimited Diffusion Works is a model based on the works of animation studio ufotable, whos works including 14 Animated Films, 15 TV Shows, and Opening Animations and Cutscenes for 17 Video Game works for several different franchises, giving a diverse palate of different animators, styles, settings, and subjects while having a recognizable visual finish due to their CGI usage and consistent color grading technique.
|
| 21 |
|
| 22 |
This model’s current version was trained on 09/2023, with around 200k images over NovelAI v1, uses the following content from studio ufotable:
|
| 23 |
+
-Fate Zero (TV)
|
| 24 |
+
-The Garden of Sinners 8 – Epilogue (Movie)
|
| 25 |
+
-The Garden of Sinners 9 - Future Gospel (Movie)
|
| 26 |
+
-The Garden of Sinners 9.5 - Future Gospel: Extra Chorus (Movie – 31m 53s)
|
| 27 |
+
-Fate Stay Night [Réalta Nua] (VG – 4m 30s)
|
| 28 |
+
-Fate Stay Night Unlimited Blade Works (TV)
|
| 29 |
+
-Fate/Hollow Ataraxia (VG – 3m)
|
| 30 |
+
-Fate Stay Night Heaven’s Feel Trilogy (Movies – 5h 59m 13s)
|
| 31 |
|
| 32 |
In addition to ufotable, I added for experimentation purposes the dataset from the following non ufotable works for additional data and regularization purposes:
|
| 33 |
+
-Lord El-Melloi II Case Files (TV – Studio Troyca)
|
| 34 |
+
-*and additional miscellaneous non-animation frame data not from danbooru*
|
| 35 |
|
| 36 |
Dataset was created by running Blu-ray rips of the listed content through ffmpeg as part of the aid of Anime-Screenshot-Pipeline on Github back in February of 2023, but I had quickly become dissatisfied with the quality of the dataset it originally produced, and parts of its pipeline were not useful for what I needed, thus I took some tools, repurposed and/or modified settings, and redid the entire dataset after several months of doing initial training tests before committing to an overall.
|
| 37 |
|