Update T5 attention mask handling in FLUX

This commit is contained in:
Kohya S
2024-08-21 08:02:33 +09:00
parent 6ab48b09d8
commit 7e459c00b2
7 changed files with 101 additions and 50 deletions

View File

@@ -9,6 +9,9 @@ __Please update PyTorch to 2.4.0. We have tested with `torch==2.4.0` and `torchv
The command to install PyTorch is as follows:
`pip3 install torch==2.4.0 torchvision==0.19.0 --index-url https://download.pytorch.org/whl/cu124`
Aug 21, 2024:
The specification of `--apply_t5_attn_mask` has been changed. Previously, the T5 output was zero-padded, but now, two steps are taken: "1. Apply mask when encoding T5" and "2. Apply mask in the attention of Double Block". Fine tuning, LoRA training, and inference in `flux_mini_inference.py` have been changed.
Aug 20, 2024 (update 3):
__Experimental__ The multi-resolution training is now supported with caching latents to disk.