Welcome to my custom-trained sd15 model.

This model was trained using around 6,000 generated flux images, experimenting with various styles (and plenty of trial and error).

Dataset from:

https://civarchive.com/models/631007?modelVersionId=705402

Since I relied on the pharmapsychotic -> clip_interrogator for automated captions, there aren’t any strict tags or descriptions to follow—just a playful, exploratory approach.

I’m doing this purely as a hobby, without previous machine learning experience.

Between the tinkering, the hours spent using onetrainer, and all the wasted energy, it’s been quite an adventure. But hey, there’s always more to learn, and I look forward to pushing this model further.

At the moment faces on this model look bad or not Idk. Maybe some good prompters can tell me. Gotta learn much more. One Knob within comfyui helped tremendously for "better prompts"

Best is to use some HiRes scaling for images from 512x512 to 768x768 for better quality.

Training Settings:

Well,

Training Configuration

• Epochs: 100

• Batch Size: On first train 10 -> 32

• Gradient Accumulation Steps: 1

• Learning Rate: 1e-05

• Weight DType: FLOAT_16

• Output DType: FLOAT_16

• Scheduler: CONSTANT

• Train Device: cuda

• Train DType: BFLOAT_16 (with fallback to FLOAT_32)

• Attention Mechanism: SDP

• Resolution: 512

Validation & Sampling

• Validation Enabled: true

• Validate After: 1 (EPOCH)

• Sample After: 200 (STEP)

• Sample Image Format: PNG

Optimizer

• Optimizer: ADAMW_8BIT

• Weight Decay: 0.01

• Beta1: 0.9

• Beta2: 0.999

• 8-bit Block-Wise: true

• Stochastic Rounding: true

It was first trained on 1e-5 then reduced to 5e-5 and so on.