kayfahaarukku / AkashicPulse-v1.0

AkashicPulse v1.0

AkashicPulse is a finetune based on RouWei, an Illustrious-based model.

The model has gone through 1 step of merging, and 3 steps of finetuning to make sure the model able to give stunning results, superior from the competitions.

Recommended settings:

• Sampling: Euler a

• Steps: 20-30, the sweet spot is 28.

• CFG: 4-10, the sweet spot is 7.

• [Not mandatory] On reForge or ComfyUI, have MaHiRo CFG enabled.

Recommended prompting format:

• Prompt: [1girl/1boy], [character name], [series], by [artist name], [the rest of the prompt], masterpiece, best quality

• Negative prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, signature, watermark, username, blurry, [the rest of the negative prompt]

Training Process:

• Step 1:

  • Giving RouWei a CyberFix treatment.

• Step 2:

  • Training new concept

    • Dataset size: ~10.000 images

    • GPU: 2xA100 80GB

    • Optimizer: AdaFactor

    • Unet Learning Rate: 7.5e-6

    • Text Encoder Learning Rate: 3.75e-6

    • Batch Size: 16

    • Gradient Accumulation: 3

    • Warmup steps: 2 * 100 steps

    • Min SNR: 5

    • Epoch: 10

    • Random Cropping: True

    • Loss: Huber

    • Huber Schedule: SNR

• Step 3:

  • Finetuning I

    • Dataset size: ~4.500 images

    • GPU: 1xA100 80GB

    • Optimizer: AdaFactor

    • Unet Learning Rate: 3e-6

    • Text Encoder Learning Rate: N/A

    • Batch Size: 16

    • Gradient Accumulation: 3

    • Warmup steps: 5%

    • Min SNR: 5

    • Epoch: 15

    • Random Cropping: True

    • Loss: Huber

    • Huber Schedule: SNR

    • Multires Noise Iteration: 8

• Step 4:

  • Finetuning II

    • Dataset size: ~4.500 images

    • GPU: 1xA100 80GB

    • Optimizer: AdaFactor

    • Unet Learning Rate: 3e-6

    • Text Encoder Learning Rate: N/A

    • Batch Size: 48

    • Gradient Accumulation: 1

    • Warmup steps: 5%

    • Min SNR: 5

    • Epoch: 15

    • Loss: L2

    • Noise Offset: 0.0357

Added series:

• DanDaDan

The model falls under Fair AI Public License 1.0-SD with no additional terms.