VictorMylle/Kohya-ss-sd-scripts
1
0
Fork 1
mirror of https://github.com/kohya-ss/sd-scripts.git synced 2026-04-09 06:45:09 +00:00
Code Issues Packages Projects Releases Wiki Activity

update README, remove unnecessary autocast

Browse Source
This commit is contained in:
kohya-ss
2024-10-19 19:21:12 +09:00
parent ef70aa7b42
commit 2c45d979e6
2 changed files with 5 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
README.md
View File

@@ -13,13 +13,13 @@ The command to install PyTorch is as follows:
Oct 19, 2024:
- Added an implementation of Differential Output Preservation (temporary name) for SDXL/FLUX.1 LoRA training.
- Added an implementation of Differential Output Preservation (temporary name) for SDXL/FLUX.1 LoRA training. SD1/2 is not tested yet. This is an experimental feature.
- A method to make the output of LoRA closer to the output when LoRA is not applied, with captions that do not contain trigger words.
- Define a Dataset subset for the regularization image (`is_reg = true`) with `.toml`. Add `custom_attributes.diff_output_preservation = true`.
- See [dataset configuration](docs/config_README-en.md) for the regularization dataset.
- Specify "number of training images x number of epochs >= number of regularization images x number of epochs".
- Specify a large value for `--prior_loss_weight` option (not dataset config). We recommend 10-1000.
- Set the loss in the training without using the regularization image to be close to the loss in the training using DOP.
- Specify "number of training images x number of repeats >= number of regularization images x number of repeats".
- Specify a large value for `--prior_loss_weight` option (not dataset config). The appropriate value is unknown, but try around 10-100. Note that the default is 1.0.
- You may want to start with 2/3 to 3/4 of the loss value when DOP is not applied. If it is 1/2, DOP may not be working.
```
[[datasets.subsets]]
image_dir = "path/to/image/dir"
@@ -28,8 +28,6 @@ is_reg = true
custom_attributes.diff_output_preservation = true # Add this
```
Oct 13, 2024:
- Fixed an issue where it took a long time to load the image size when initializing the dataset, especially when the number of images in the dataset was large.

2
flux_train_network.py
View File

@@ -453,7 +453,7 @@ class FluxNetworkTrainer(train_network.NetworkTrainer):
if len(diff_output_pr_indices) > 0:
network.set_multiplier(0.0)
with torch.no_grad(), accelerator.autocast():
with torch.no_grad():
model_pred_prior = call_dit(
img=packed_noisy_model_input[diff_output_pr_indices],
img_ids=img_ids[diff_output_pr_indices],