bmaltais

Can someone share a toml config file for a simple one concept finetuning? I never do finetuning and apparently using .toml is the way to go now... and I have...

I have figured it out... ``` [[datasets]] resolution = 1024 batch_size = 4 keep_tokens = 1 enable_bucket = true [[datasets.subsets]] image_dir = 'd:\kohya_ss\examples\stable_cascade\test_dataset' num_repeats = 10 class_tokens = 'toy' caption_extension...

Look like I am successful in finetuning...  Finetuning as `zxc` class `toy` and prompting with `zxc toy posing at the beach --W 800 --H 1200`... so there is hope...

I have shared the test dataset in the stable_cascade branch. Look under the examples folder. You can play with it for now.

I tested the results of th model in COmfyUI and they are not great... sort of washed out... Most certainly bad training parameters... Will take a while to figureout proper...

If you find better parameters for better results please share. Training SC is hugely VRAM intensive.

I did my test with 8… I don’t think the disappointing result is due to that… I tried using other optimiser but I don’t have enough vram.

Using the latest updated code in sd-scripts produce better results... still not perfect... kohya is working on allowing to train stage_b... hoping this will fix the issue with the final...

I did provide everything in the stable_cascade branch. Look in the example folder in that branch. You will find the dataset, toml file for the dataset, etc. The new way...

The dataset can be anywhere. Simply edit the toml file to point to it and specify the repeats, resolution, etc.