Support me on Ko-fi:
View the announcements on my profile for information regarding requests/commissions.
Request
Moss Version:
Takash-ish aka Crapeuchi Crapashi. Not really fully happy with how it came out but this was mostly an experiment. I may continue with Moss in the future but with how long the training takes and amount of errors I receive in gens it seems unlikely.
Works best on moss_image_sft_int8_convrot as that's what it was trained on (tried on bf16 but PC couldn't handle the drip).
sd-scripts moss_image_train_network.py:
(LR and/or epochs likely needs raised but I didn't want to invest in another 8 hour training session)
Epochs: 10
LR: 2e-5, dim/alpha 32/32
Steps: 4800 with batch 1 and gradient accumulation 1
optimizer: optimi.AdamW, betas=(0.9, 0.99), eps=1e-8, weight_decay=0.01
scheduler: cosine
Trained subsets at mixed resolution 512/1024/1536.
Mixed captioning combining Danbooru tags and natural language