Is it possible to train 512x 768 images?
Closed this issue · 9 comments
I saw other Lora Trainer has that options, but they required xformers which cannot be ran in AMD GPU, is it possible to train no squad size image in this trainer?
You can train with any size or resolution thanks to buckets which are turned on by default
You can train with any size or resolution thanks to buckets which are turned on by default
what resolution scale should I set, greater than the maximum dimension of the largest photo?
You should train at 512 resolution for speed, or go higher like 768 if you need small details or even 1024 for stuff like unique pupils. The training resolution is irrelevant and only affects the precision of smaller details.
Many thx
Btw how to datasets.subsets muiltiple datasets.subsets folders into one datasets and train now, I still don't very sure the format of dataset_config.toml. It showed " 💥 Error: Your custom dataset is invalid or contains an error! Please check the original template." when I tried to run the trainer
Give the original example to chat gpt and ask it to make another one for your case.
Give the original example to chat gpt and ask it to make another one for your case.
May I ask what question should I ask Chat GPT to get the right answer? I am so sorry I have just trying out SD for few days
This gives me the correct answer. Personally I understand the initial format just as well as the description with natural language, but Chat GPT helps a lot if you're not as computer savvy.
Here's a format for configuring folders in toml:
[[datasets]] [[datasets.subsets]] image_dir = "/content/drive/MyDrive/Loras/example/dataset/good_images" num_repeats = 3 [[datasets.subsets]] image_dir = "/content/drive/MyDrive/Loras/example/dataset/normal_images" num_repeats = 1Please produce a configuration in the same format where instead of "example" my project is called "cool_lora", and instead of good and bad images I have: "folder1" with 1 repeat, "folder2" with 2 repeats, "folder3" with 5 repeats
Answer:
Sure thing! Here's the configuration you requested:
[[datasets]] [[datasets.subsets]] image_dir = "/content/drive/MyDrive/Loras/cool_lora/dataset/folder1" num_repeats = 1 [[datasets.subsets]] image_dir = "/content/drive/MyDrive/Loras/cool_lora/dataset/folder2" num_repeats = 2 [[datasets.subsets]] image_dir = "/content/drive/MyDrive/Loras/cool_lora/dataset/folder3" num_repeats = 5Let me know if you need any other help with your project!
many thx