It's seems that v0.4.8 produce more relevant images with same default hp than v0.4.9, v0.4.10.

[ZackPashkin]

Hi! It's seems that v0.4.8 produce more relevant images with same default hp than v0.4.9, v0.4.10.
Also predictions looks same, didn't change much, after first 30-100 iterations.
Any ideas how it can be improved

[DrJKL]

Hey Zack,

Do you have any examples?

477b832...main

There shouldn't be any substantive differences in the generation between those versions.

[ZackPashkin]

Hi @DrJKL Here is some examples to compare: promt is an apple next to a fireplace Now seems 0.1.1 produced more relevant results on 1 epoch, comparing to 0.4.8, 0.4.9, 0.4.10
LEARNING_RATE = 5e-2
SEED = 0

[ZackPashkin]

Seems ~ lr from 13e-2 to 20e-2 , iterations = 200 produce more semantically relevant results. What do you observe?
Also what do you think about prompt engineering . How long should be optimal length and should noun and adjective be duplicative

[LtqxWYEG]

I found that lr values over 0.11 appear to produce more non-converging garbage than values below. (But that is HEAVILY dependent on the seed) Also there are significant changes in the image up to 15k+ iterations. Just that they take longer to emerge