UrangDiffusion v2.0

[UrangDiffusion v2.0 finetuning is sponsored by CagliostroLab, the makers of Animagine XL]

UrangDiffusion v2.0 (oo-raw-ng Diffusion) brings a whole-new training method compared to the v1.4. The model provide more flexibility and brings some updated dataset.

The name “Urang” comes from Sundanese, meaning “We/Our/I.” The history behind the name is to make the model not only suitable for me but also for many people. Another reason is that I use many resources (training scripts, dataset collecting scripts, etc.) from other people. It’s unfair to claim this model as “my sole work.”

Standard Prompting Guidelines

The model is finetuned from Animagine XL 3.1, which is trained with danbooru tags. However, there is a little bit changes on dataset captioning, therefore there is some different default prompt used:

• Default prompt: 1girl/1boy, character name, from what series, everything else in any order, best quality, amazing quality, very aesthetic.

• Note: The quality tag masterpiece has been replaced with best quality due to reports that it often caused unwanted side effects. Tests also proven that the anatomy of some generations are broken because of the tag.

• Default negative prompt: lowres, (bad), text, error, fewer, extra, missing, worst quality, jpeg artifacts, low quality, watermark, unfinished, displeasing, oldest, early, chromatic aberration, signature, extra digits, artistic error, username, scan, [abstract],

• Default configuration: Euler a with around 25-30 steps, CFG 5-7, and ENSD set to 31337. Sweet spot is around 28 steps and CFG 7.

Training Configurations

Finetuned from: Animagine XL 3.1

Pretraining:

• Dataset size: 44,393 images

• GPU: 1xA100 80GB

• Optimizer: AdaFactor

• Unet Learning Rate: 3.75e-6

• Text Encoder Learning Rate: 1.875e-6

• Batch Size: 16

• Gradient Accumulation: 3

• Warmup steps: 100 steps

• Min SNR: 5

• Epoch: 10

• Random Cropping: True

• Loss: Huber

• Huber Schedule: SNR

• Huber C: 0.1

Finetuning:

• Dataset size: 3,140 images

• GPU: 1xA100 80GB

• Optimizer: AdaFactor

• Unet Learning Rate: 3e-6

• Text Encoder Learning Rate: - (Train TE set to False)

• Batch Size: 16

• Gradient Accumulation: 3

• Warmup steps: 5%

• Min SNR: 5

• Epoch: 10 (epoch 9 is used)

• Noise Offset: 0.0357

• Random Cropping: True

• Loss: Huber

• Huber Schedule: SNR

• Huber C: 0.1

Added/Updated Series and Characters

Series:

1. zenless zone zero

2. wuthering waves

3. sewayaki kitsune no senko-san

Honkai: Star Rail:

1. firefly

2. acheron

3. sparkle

4. robin

5. aventurine

6. black swan

7. feixiao

8. yunli

9. lingsha

10. march 7th (hunt)

11. jade

12. jiaoqiu

13. gallagher

14. rappa

15. misha

Hololive Talents:

1. hololive indonesia

2. raora panthera

3. elizabeth rose bloodflame

4. gigi murin

5. cecilia immergreen

Genshin Impact:

1. arlecchino

2. clorinde

3. chiori

4. mualani

5. xianyun

6. sigewinne

7. kinich

8. xilonen

9. emilie

10. gaming

11. kachina

12. sethos

Others:

1. landscape

2. several concepts to fix anatomy issue

Special Thanks

• CagliostroLab for sponsoring the model finetuning by letting me borrowed the organization’s RunPod account.

• My co-workers(?) at CagliostroLab for the insights and feedback.

• Nur Hikari and Vanilla Latte for quality control.

• Linaqruf, my tutor and role model in AI-generated images, and also the person behind tag ordering.

License

UrangDiffusion falls under the Fair AI Public License 1.0-SD license.