theluyuan/DiffSynth-Studio
1
0
Fork 0
mirror of https://github.com/modelscope/DiffSynth-Studio.git synced 2026-03-21 16:18:13 +00:00
Code Issues Packages Projects Releases Wiki Activity

Flux lora update (#237)

Browse Source 
* update flux lora

---------

Co-authored-by: tc2000731 <tc2000731@163.com>
This commit is contained in:
Zhongjie Duan
2024-10-11 18:41:24 +08:00
committed by GitHub
parent 75ab786afc
commit 22e4ae99e8
11 changed files with 63 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
examples/train/README.md
View File

@@ -123,7 +123,7 @@ models/FLUX/
└── model.safetensors.index.json
```
Launch the training task using the following command:
Launch the training task using the following command (39G VRAM required):
```
CUDA_VISIBLE_DEVICES="0" python examples/train/flux/train_flux_lora.py \
@@ -134,18 +134,20 @@ CUDA_VISIBLE_DEVICES="0" python examples/train/flux/train_flux_lora.py \
--dataset_path data/dog \
--output_path ./models \
--max_epochs 1 \
--steps_per_epoch 500 \
--steps_per_epoch 100 \
--height 1024 \
--width 1024 \
--center_crop \
--precision "bf16" \
--learning_rate 1e-4 \
--lora_rank 4 \
--lora_alpha 4 \
--lora_rank 16 \
--lora_alpha 16 \
--use_gradient_checkpointing \
--align_to_opensource_format
```
By adding parameter `--quantize "float8_e4m3fn"`, you can save approximate 10G VRAM.
**`--align_to_opensource_format` means that this script will export the LoRA weights in the opensource format. This format can be loaded in both DiffSynth-Studio and other codebases.**
For more information about the parameters, please use `python examples/train/flux/train_flux_lora.py -h` to see the details.