UrangDiffusion | an AingDiffusion XL sequel

https://civitai.com/models/537384?modelVersionId=690340

UrangDiffusion v1.3 (oo-raw-ng Diffusion) is an updated version of UrangDiffusion 1.2. This version provides refreshed dataset, improvements over the last iteration, training parameter correction, and some characters optimization.

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.”

The model went through two steps of training: pretraining and finetuning. Pretraining is to make the model learn new things, while finetuning ensures the images produced by the model are decent (A.K.A. having a standard style) without mentioning style in the prompt.

Standard Prompting Guidelines The model is finetuned from Animagine XL 3.1. 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, masterpiece, best quality, amazing quality, very aesthetic

Default negative prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name, displeasing

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: 27,545 images

GPU: 1xA100 80GB

Optimizer: AdaFactor

Unet Learning Rate: 3.75e-6

Text Encoder Learning Rate: 1.875e-6

Batch Size: 48

Gradient Accumulation: 1

Warmup steps: 100 steps

Min SNR: 5

Epoch: 10

Finetuning:

Dataset size: ~6,800 images

GPU: 1xA100 80GB

Optimizer: AdaFactor

Unet Learning Rate: 2e-6

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

Batch Size: 48

Gradient Accumulation: 1

Warmup steps: 5%

Min SNR: 5

Epoch: 10

Noise Offset: 0.0357

Added Series Wuthering Waves, Zenless Zone Zero and hololiveEN -Justice- have been added to the model.

Special Thanks 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.

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