Model pipeline

The product trained identity once, then branched into selected styles.

The GPU server used my own DreamBooth implementation to train a personalized checkpoint from the uploaded photos. Custom PyTorch code then merged that identity checkpoint into each selected style model. Reusing the checkpoint avoided training once per style and saved substantial GPU time. GPU compute was by far the product's largest expense.

• Identity once: one person checkpoint captures the uploaded subject.
• Style branches: selected package themes decide which custom style checkpoints get merged.
• Prompt batches: sex, age, hair, eyes, and glasses fill themed prompt templates.

model tree · package-selected branches

Uploaded photosvalidated 512x512 face crops from the browser

↓

Person identity checkpointtrained once per order from customer photos

↓

cartoon-base-> Cartoon portraits

pixar-base-> American portraits

ghibli-base-> Ghibli portraits

shinkai-base-> Shinkai portraits

Software gives me a fast, mostly deterministic feedback loop. Model training does not. The hardest part of this project was ensuring good results. I made my own DreamBooth implementation. I also customized training for each customer: I selected class-specific regularization data based on the customer’s subject class and normalized epoch count against the number of uploaded photos to target approximately 1,700 total training steps. I built prompt templates around characteristics such as hair, eyes, and age. I tested sampler settings and seeds, spent weeks curating better training sets, and added an upscaler step when customers paid for larger outputs. A major breakthrough was model merging. Instead of training each customer in four different styles, each with a different base model, I trained the identity once and transferred its learned delta into each style checkpoint. Model merging was still novel then, but it cut GPU time in half on average while preserving quality in the final outputs.