Why?
This LoRA model aims to solve the "strawberry" problem of ML image generation models.
"Oh, your billion dollar algorithm can not solve the full wine glass problem!"
Not anymore.
Open Source is For Everyone
Open source is not bound by huge corporate workflows and processes. It took me 32 minutes to train this model after manually captioning 20 images.
Training Details:
Epochs: 12
Steps: 1920
Optimizer: --optimizer_type=adopt.ADOPT
LR: 8e-5
TE LR: 4e-5
Scheduler: constant_with_warmup 2% (important to warm the refrigerated wine)
Rank: 128/64
Debiased Est. Loss: True
No flipping or caption shuffling because flipping would not work well with clocks (duh).
Problems Encountered During Preparation & Training:
It's slightly harder to find full wine glass pictures and clocks showing different time (duh). Thanks reddit for the clocks and instagram for full wine glasses. Some people are animals.
Turns out I forgot how to read analog clocks. Some images might have been captioned wrong.
The wine and clock biases are harder to fully correct than I expected.
OOD generations still underperform. More examples and steps might be needed.
Description
Trained with bf16 mixed precision and saved as float to increase accuracy.