Finetuning realvisXL
#5
by
KeilahElla
- opened
This is an incredibly good model, many thanks for your efforts!
I am trying to finetune a checkpoint using this model as seed (using Kohya). Do you think the following parameters make sense?
"optimizer": "Adafactor",
"optimizer_args": "scale_parameter=False relative_step=False warmup_init=False",
"learning_rate": 1e-06,
"learning_rate_te1": 1e-06,
"learning_rate_te2": 1e-06,
"max_token_length": "150",
"additional_parameters": "--max_grad_norm=0.0 --no_half_vae --train_text_encoder",
If I go for a larger learning rate (>5e-6), then the model starts to output very blurry images, especially if the negative prompts are left blank. I think the model starts learning jpeg compression artefacts.