theluyuan/DiffSynth-Studio
1
0
Fork 0
mirror of https://github.com/modelscope/DiffSynth-Studio.git synced 2026-03-18 22:08:13 +00:00
Code Issues Packages Projects Releases Wiki Activity

rebuild base modules

Browse Source
This commit is contained in:
Artiprocher
2024-07-26 12:15:40 +08:00
parent 9471bff8a4
commit e3f8a576cf
76 changed files with 3253 additions and 3563 deletions
Download Patch File Download Diff File Expand all files Collapse all files

379
examples/train/README.md Normal file
View File

@@ -0,0 +1,379 @@
# DiffSynth Training Framework
We have implemented a training framework for text-to-image Diffusion models, enabling users to easily train LoRA models using our framework. Our provided scripts come with the following advantages:
* **Comprehensive Functionality & User-Friendliness**: Our training framework supports multi-GPU and multi-machine setups, facilitates the use of DeepSpeed for acceleration, and includes gradient checkpointing optimizations for models with excessive memory demands.
* **Code Conciseness & Researcher Accessibility**: We avoid large blocks of complicated code. General-purpose modules are implemented in `diffsynth/trainers/text_to_image.py`, while model-specific training scripts contain only minimal code pertinent to the model architecture, making it researcher-friendly.
* **Modular Design & Developer Flexibility**: Built on the universal Pytorch-Lightning framework, our training framework is decoupled in terms of functionality, allowing developers to easily introduce additional training techniques by modifying our scripts to suit their needs.
Image Examples of fine-tuned LoRA. The prompt is "一只小狗蹦蹦跳跳，周围是姹紫嫣红的鲜花，远处是山脉" (for Chinese models) or "a dog is jumping, flowers around the dog, the background is mountains and clouds" (for English models).
||Kolors|Stable Diffusion 3|Hunyuan-DiT|
|-|-|-|-|
|Without LoRA|![image_without_lora](https://github.com/modelscope/DiffSynth-Studio/assets/35051019/9d79ed7a-e8cf-4d98-800a-f182809db318)|![image_without_lora](https://github.com/modelscope/DiffSynth-Studio/assets/35051019/ddb834a5-6366-412b-93dc-6d957230d66e)|![image_without_lora](https://github.com/Artiprocher/DiffSynth-Studio/assets/35051019/1aa21de5-a992-4b66-b14f-caa44e08876e)|
|With LoRA|![image_with_lora](https://github.com/modelscope/DiffSynth-Studio/assets/35051019/02f62323-6ee5-4788-97a1-549732dbe4f0)|![image_with_lora](https://github.com/modelscope/DiffSynth-Studio/assets/35051019/8e7b2888-d874-4da4-a75b-11b6b214b9bf)|![image_with_lora](https://github.com/Artiprocher/DiffSynth-Studio/assets/35051019/83a0a41a-691f-4610-8e7b-d8e17c50a282)|
## Prepare your dataset
We provide an example dataset [here](https://modelscope.cn/datasets/buptwq/lora-stable-diffusion-finetune/files). You need to manage the training images as follows:
```
data/dog/
└── train
├── 00.jpg
├── 01.jpg
├── 02.jpg
├── 03.jpg
├── 04.jpg
└── metadata.csv
```
`metadata.csv`:
```
file_name,text
00.jpg,a dog
01.jpg,a dog
02.jpg,a dog
03.jpg,a dog
04.jpg,a dog
```
Note that if the model is Chinese model (for example, Hunyuan-DiT and Kolors), we recommand to use Chinese texts in the dataset. For example
```
file_name,text
00.jpg,一只小狗
01.jpg,一只小狗
02.jpg,一只小狗
03.jpg,一只小狗
04.jpg,一只小狗
```
## Train a LoRA model
### Kolors
The following files will be used for constructing Kolors. You can download Kolors from [huggingface](https://huggingface.co/Kwai-Kolors/Kolors) or [modelscope](https://modelscope.cn/models/Kwai-Kolors/Kolors). Due to precision overflow issues, we need to download an additional VAE model (from [huggingface](https://huggingface.co/madebyollin/sdxl-vae-fp16-fix) or [modelscope](https://modelscope.cn/models/AI-ModelScope/sdxl-vae-fp16-fix)). You can use the following code to download these files:
```python
from diffsynth import download_models
download_models(["Kolors", "SDXL-vae-fp16-fix"])
```
```
models
├── kolors
│ └── Kolors
│ ├── text_encoder
│ │ ├── config.json
│ │ ├── pytorch_model-00001-of-00007.bin
│ │ ├── pytorch_model-00002-of-00007.bin
│ │ ├── pytorch_model-00003-of-00007.bin
│ │ ├── pytorch_model-00004-of-00007.bin
│ │ ├── pytorch_model-00005-of-00007.bin
│ │ ├── pytorch_model-00006-of-00007.bin
│ │ ├── pytorch_model-00007-of-00007.bin
│ │ └── pytorch_model.bin.index.json
│ ├── unet
│ │ └── diffusion_pytorch_model.safetensors
│ └── vae
│ └── diffusion_pytorch_model.safetensors
└── sdxl-vae-fp16-fix
└── diffusion_pytorch_model.safetensors
```
Launch the training task using the following command:
```
CUDA_VISIBLE_DEVICES="0" python examples/train/kolors/train_kolors_lora.py \
--pretrained_unet_path models/kolors/Kolors/unet/diffusion_pytorch_model.safetensors \
--pretrained_text_encoder_path models/kolors/Kolors/text_encoder \
--pretrained_fp16_vae_path models/sdxl-vae-fp16-fix/diffusion_pytorch_model.safetensors \
--dataset_path data/dog \
--output_path ./models \
--max_epochs 1 \
--steps_per_epoch 500 \
--height 1024 \
--width 1024 \
--center_crop \
--precision "16-mixed" \
--learning_rate 1e-4 \
--lora_rank 4 \
--lora_alpha 4 \
--use_gradient_checkpointing
```
For more information about the parameters, please use `python examples/train/kolors/train_kolors_lora.py -h` to see the details.
After training, use `model_manager.load_lora` to load the LoRA for inference.
```python
from diffsynth import ModelManager, SDXLImagePipeline
import torch
model_manager = ModelManager(torch_dtype=torch.float16, device="cuda",
file_path_list=[
"models/kolors/Kolors/text_encoder",
"models/kolors/Kolors/unet/diffusion_pytorch_model.safetensors",
"models/sdxl-vae-fp16-fix/diffusion_pytorch_model.safetensors"
])
model_manager.load_lora("models/lightning_logs/version_0/checkpoints/epoch=0-step=500.ckpt", lora_alpha=1.0)
pipe = SDXLImagePipeline.from_model_manager(model_manager)
torch.manual_seed(0)
image = pipe(
prompt="一只小狗蹦蹦跳跳，周围是姹紫嫣红的鲜花，远处是山脉",
negative_prompt="",
cfg_scale=7.5,
num_inference_steps=100, width=1024, height=1024,
)
image.save("image_with_lora.jpg")
```
### Stable Diffusion 3
Only one file is required in the training script. You can use [`sd3_medium_incl_clips.safetensors`](https://huggingface.co/stabilityai/stable-diffusion-3-medium/resolve/main/sd3_medium_incl_clips.safetensors) (without T5 encoder) or [`sd3_medium_incl_clips_t5xxlfp16.safetensors`](https://huggingface.co/stabilityai/stable-diffusion-3-medium/resolve/main/sd3_medium_incl_clips_t5xxlfp16.safetensors) (with T5 encoder). Please use the following code to download these files:
```python
from diffsynth import download_models
download_models(["StableDiffusion3", "StableDiffusion3_without_T5"])
```
```
models/stable_diffusion_3/
├── Put Stable Diffusion 3 checkpoints here.txt
├── sd3_medium_incl_clips.safetensors
└── sd3_medium_incl_clips_t5xxlfp16.safetensors
```
Launch the training task using the following command:
```
CUDA_VISIBLE_DEVICES="0" python examples/train/stable_diffusion_3/train_sd3_lora.py \
--pretrained_path models/stable_diffusion_3/sd3_medium_incl_clips.safetensors \
--dataset_path data/dog \
--output_path ./models \
--max_epochs 1 \
--steps_per_epoch 500 \
--height 1024 \
--width 1024 \
--center_crop \
--precision "16-mixed" \
--learning_rate 1e-4 \
--lora_rank 4 \
--lora_alpha 4 \
--use_gradient_checkpointing
```
For more information about the parameters, please use `python examples/train/stable_diffusion_3/train_sd3_lora.py -h` to see the details.
After training, use `model_manager.load_lora` to load the LoRA for inference.
```python
from diffsynth import ModelManager, SD3ImagePipeline
import torch
model_manager = ModelManager(torch_dtype=torch.float16, device="cuda",
file_path_list=["models/stable_diffusion_3/sd3_medium_incl_clips.safetensors"])
model_manager.load_lora("models/lightning_logs/version_0/checkpoints/epoch=0-step=500.ckpt", lora_alpha=1.0)
pipe = SD3ImagePipeline.from_model_manager(model_manager)
torch.manual_seed(0)
image = pipe(
prompt="a dog is jumping, flowers around the dog, the background is mountains and clouds",
negative_prompt="bad quality, poor quality, doll, disfigured, jpg, toy, bad anatomy, missing limbs, missing fingers, 3d, cgi, extra tails",
cfg_scale=7.5,
num_inference_steps=100, width=1024, height=1024,
)
image.save("image_with_lora.jpg")
```
### Hunyuan-DiT
Four files will be used for constructing Hunyuan DiT. You can download them from [huggingface](https://huggingface.co/Tencent-Hunyuan/HunyuanDiT) or [modelscope](https://www.modelscope.cn/models/modelscope/HunyuanDiT/summary). You can use the following code to download these files:
```python
from diffsynth import download_models
download_models(["HunyuanDiT"])
```
```
models/HunyuanDiT/
├── Put Hunyuan DiT checkpoints here.txt
└── t2i
├── clip_text_encoder
│ └── pytorch_model.bin
├── model
│ └── pytorch_model_ema.pt
├── mt5
│ └── pytorch_model.bin
└── sdxl-vae-fp16-fix
└── diffusion_pytorch_model.bin
```
Launch the training task using the following command:
```
CUDA_VISIBLE_DEVICES="0" python examples/train/hunyuan_dit/train_hunyuan_dit_lora.py \
--pretrained_path models/HunyuanDiT/t2i \
--dataset_path data/dog \
--output_path ./models \
--max_epochs 1 \
--steps_per_epoch 500 \
--height 1024 \
--width 1024 \
--center_crop \
--precision "16-mixed" \
--learning_rate 1e-4 \
--lora_rank 4 \
--lora_alpha 4 \
--use_gradient_checkpointing
```
For more information about the parameters, please use `python examples/train/hunyuan_dit/train_hunyuan_dit_lora.py -h` to see the details.
After training, use `model_manager.load_lora` to load the LoRA for inference.
```python
from diffsynth import ModelManager, HunyuanDiTImagePipeline
import torch
model_manager = ModelManager(torch_dtype=torch.float16, device="cuda",
file_path_list=[
"models/HunyuanDiT/t2i/clip_text_encoder/pytorch_model.bin",
"models/HunyuanDiT/t2i/model/pytorch_model_ema.pt",
"models/HunyuanDiT/t2i/mt5/pytorch_model.bin",
"models/HunyuanDiT/t2i/sdxl-vae-fp16-fix/diffusion_pytorch_model.bin"
])
model_manager.load_lora("models/lightning_logs/version_0/checkpoints/epoch=0-step=500.ckpt", lora_alpha=1.0)
pipe = HunyuanDiTImagePipeline.from_model_manager(model_manager)
torch.manual_seed(0)
image = pipe(
prompt="一只小狗蹦蹦跳跳，周围是姹紫嫣红的鲜花，远处是山脉",
negative_prompt="",
cfg_scale=7.5,
num_inference_steps=100, width=1024, height=1024,
)
image.save("image_with_lora.jpg")
```
### Stable Diffusion
Only one file is required in the training script. We support the mainstream checkpoints in [CivitAI](https://civitai.com/). By default, we use the base Stable Diffusion v1.5. You can download it from [huggingface](https://huggingface.co/runwayml/stable-diffusion-v1-5/resolve/main/v1-5-pruned-emaonly.safetensors) or [modelscope](https://www.modelscope.cn/models/AI-ModelScope/stable-diffusion-v1-5/resolve/master/v1-5-pruned-emaonly.safetensors). You can use the following code to download this file:
```python
from diffsynth import download_models
download_models(["StableDiffusion_v15"])
```
```
models/stable_diffusion
├── Put Stable Diffusion checkpoints here.txt
└── v1-5-pruned-emaonly.safetensors
```
Launch the training task using the following command:
```
CUDA_VISIBLE_DEVICES="0" python examples/train/stable_diffusion/train_sd_lora.py \
--pretrained_path models/stable_diffusion/v1-5-pruned-emaonly.safetensors \
--dataset_path data/dog \
--output_path ./models \
--max_epochs 1 \
--steps_per_epoch 500 \
--height 512 \
--width 512 \
--center_crop \
--precision "16-mixed" \
--learning_rate 1e-4 \
--lora_rank 4 \
--lora_alpha 4 \
--use_gradient_checkpointing
```
For more information about the parameters, please use `python examples/train/stable_diffusion/train_sd_lora.py -h` to see the details.
After training, use `model_manager.load_lora` to load the LoRA for inference.
```python
from diffsynth import ModelManager, SDImagePipeline
import torch
model_manager = ModelManager(torch_dtype=torch.float16, device="cuda",
file_path_list=["models/stable_diffusion/v1-5-pruned-emaonly.safetensors"])
model_manager.load_lora("models/lightning_logs/version_0/checkpoints/epoch=0-step=500.ckpt", lora_alpha=1.0)
pipe = SDImagePipeline.from_model_manager(model_manager)
torch.manual_seed(0)
image = pipe(
prompt="a dog is jumping, flowers around the dog, the background is mountains and clouds",
negative_prompt="bad quality, poor quality, doll, disfigured, jpg, toy, bad anatomy, missing limbs, missing fingers, 3d, cgi, extra tails",
cfg_scale=7.5,
num_inference_steps=100, width=512, height=512,
)
image.save("image_with_lora.jpg")
```
### Stable Diffusion XL
Only one file is required in the training script. We support the mainstream checkpoints in [CivitAI](https://civitai.com/). By default, we use the base Stable Diffusion XL. You can download it from [huggingface](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/resolve/main/sd_xl_base_1.0.safetensors) or [modelscope](https://www.modelscope.cn/models/AI-ModelScope/stable-diffusion-xl-base-1.0/resolve/master/sd_xl_base_1.0.safetensors). You can use the following code to download this file:
```python
from diffsynth import download_models
download_models(["StableDiffusionXL_v1"])
```
```
models/stable_diffusion_xl
├── Put Stable Diffusion XL checkpoints here.txt
└── sd_xl_base_1.0.safetensors
```
We observed that Stable Diffusion XL is not float16-safe, thus we recommand users to use float32.
```
CUDA_VISIBLE_DEVICES="0" python examples/train/stable_diffusion_xl/train_sdxl_lora.py \
--pretrained_path models/stable_diffusion_xl/sd_xl_base_1.0.safetensors \
--dataset_path data/dog \
--output_path ./models \
--max_epochs 1 \
--steps_per_epoch 500 \
--height 1024 \
--width 1024 \
--center_crop \
--precision "32" \
--learning_rate 1e-4 \
--lora_rank 4 \
--lora_alpha 4 \
--use_gradient_checkpointing
```
For more information about the parameters, please use `python examples/train/stable_diffusion_xl/train_sdxl_lora.py -h` to see the details.
After training, use `model_manager.load_lora` to load the LoRA for inference.
```python
from diffsynth import ModelManager, SDXLImagePipeline
import torch
model_manager = ModelManager(torch_dtype=torch.float16, device="cuda",
file_path_list=["models/stable_diffusion_xl/sd_xl_base_1.0.safetensors"])
model_manager.load_lora("models/lightning_logs/version_0/checkpoints/epoch=0-step=500.ckpt", lora_alpha=1.0)
pipe = SDXLImagePipeline.from_model_manager(model_manager)
torch.manual_seed(0)
image = pipe(
prompt="a dog is jumping, flowers around the dog, the background is mountains and clouds",
negative_prompt="bad quality, poor quality, doll, disfigured, jpg, toy, bad anatomy, missing limbs, missing fingers, 3d, cgi, extra tails",
cfg_scale=7.5,
num_inference_steps=100, width=1024, height=1024,
)
image.save("image_with_lora.jpg")
```

181
examples/train/hunyuan_dit/README.md
View File

@@ -1,181 +0,0 @@
# Hunyuan DiT
Hunyuan DiT is an image generation model based on DiT. We provide training and inference support for Hunyuan DiT.
## Download models
Four files will be used for constructing Hunyuan DiT. You can download them from [huggingface](https://huggingface.co/Tencent-Hunyuan/HunyuanDiT) or [modelscope](https://www.modelscope.cn/models/modelscope/HunyuanDiT/summary).
```
models/HunyuanDiT/
├── Put Hunyuan DiT checkpoints here.txt
└── t2i
├── clip_text_encoder
│ └── pytorch_model.bin
├── model
│ └── pytorch_model_ema.pt
├── mt5
│ └── pytorch_model.bin
└── sdxl-vae-fp16-fix
└── diffusion_pytorch_model.bin
```
You can use the following code to download these files:
```python
from diffsynth import download_models
download_models(["HunyuanDiT"])
```
## Train
### Install training dependency
```
pip install peft lightning pandas torchvision
```
### Prepare your dataset
We provide an example dataset [here](https://modelscope.cn/datasets/buptwq/lora-stable-diffusion-finetune/files). You need to manage the training images as follows:
```
data/dog/
└── train
├── 00.jpg
├── 01.jpg
├── 02.jpg
├── 03.jpg
├── 04.jpg
└── metadata.csv
```
`metadata.csv`:
```
file_name,text
00.jpg,一只小狗
01.jpg,一只小狗
02.jpg,一只小狗
03.jpg,一只小狗
04.jpg,一只小狗
```
### Train a LoRA model
We provide a training script `train_hunyuan_dit_lora.py`. Before you run this training script, please copy it to the root directory of this project.
If GPU memory >= 24GB, we recommmand to use the following settings.
```
CUDA_VISIBLE_DEVICES="0" python train_hunyuan_dit_lora.py \
--pretrained_path models/HunyuanDiT/t2i \
--dataset_path data/dog \
--output_path ./models \
--max_epochs 1 \
--center_crop
```
If 12GB <= GPU memory <= 24GB, we recommand to enable gradient checkpointing.
```
CUDA_VISIBLE_DEVICES="0" python train_hunyuan_dit_lora.py \
--pretrained_path models/HunyuanDiT/t2i \
--dataset_path data/dog \
--output_path ./models \
--max_epochs 1 \
--center_crop \
--use_gradient_checkpointing
```
Optional arguments:
```
-h, --help show this help message and exit
--pretrained_path PRETRAINED_PATH
Path to pretrained model. For example, `./HunyuanDiT/t2i`.
--dataset_path DATASET_PATH
The path of the Dataset.
--output_path OUTPUT_PATH
Path to save the model.
--steps_per_epoch STEPS_PER_EPOCH
Number of steps per epoch.
--height HEIGHT Image height.
--width WIDTH Image width.
--center_crop Whether to center crop the input images to the resolution. If not set, the images will be randomly cropped. The images will be resized to the resolution first before cropping.
--random_flip Whether to randomly flip images horizontally
--batch_size BATCH_SIZE
Batch size (per device) for the training dataloader.
--dataloader_num_workers DATALOADER_NUM_WORKERS
Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process.
--precision {32,16,16-mixed}
Training precision
--learning_rate LEARNING_RATE
Learning rate.
--lora_rank LORA_RANK
The dimension of the LoRA update matrices.
--lora_alpha LORA_ALPHA
The weight of the LoRA update matrices.
--use_gradient_checkpointing
Whether to use gradient checkpointing.
--accumulate_grad_batches ACCUMULATE_GRAD_BATCHES
The number of batches in gradient accumulation.
--training_strategy {auto,deepspeed_stage_1,deepspeed_stage_2,deepspeed_stage_3}
Training strategy
--max_epochs MAX_EPOCHS
Number of epochs.
```
### Inference with your own LoRA model
After training, you can use your own LoRA model to generate new images. Here are some examples.
```python
from diffsynth import ModelManager, HunyuanDiTImagePipeline
from peft import LoraConfig, inject_adapter_in_model
import torch
def load_lora(dit, lora_rank, lora_alpha, lora_path):
lora_config = LoraConfig(
r=lora_rank,
lora_alpha=lora_alpha,
init_lora_weights="gaussian",
target_modules=["to_q", "to_k", "to_v", "to_out"],
)
dit = inject_adapter_in_model(lora_config, dit)
state_dict = torch.load(lora_path, map_location="cpu")
dit.load_state_dict(state_dict, strict=False)
return dit
# Load models
model_manager = ModelManager(torch_dtype=torch.float16, device="cuda")
model_manager.load_models([
"models/HunyuanDiT/t2i/clip_text_encoder/pytorch_model.bin",
"models/HunyuanDiT/t2i/mt5/pytorch_model.bin",
"models/HunyuanDiT/t2i/model/pytorch_model_ema.pt",
"models/HunyuanDiT/t2i/sdxl-vae-fp16-fix/diffusion_pytorch_model.bin"
])
pipe = HunyuanDiTImagePipeline.from_model_manager(model_manager)
# Generate an image with lora
pipe.dit = load_lora(
pipe.dit,
lora_rank=4, lora_alpha=4.0, # The two parameters should be consistent with those in your training script.
lora_path="path/to/your/lora/model/lightning_logs/version_x/checkpoints/epoch=x-step=xxx.ckpt"
)
torch.manual_seed(0)
image = pipe(
prompt="一只小狗蹦蹦跳跳，周围是姹紫嫣红的鲜花，远处是山脉",
negative_prompt="",
num_inference_steps=50, height=1024, width=1024,
)
image.save("image_with_lora.png")
```
Prompt: 一只小狗蹦蹦跳跳，周围是姹紫嫣红的鲜花，远处是山脉
|Without LoRA|With LoRA|
|-|-|
|![image_without_lora](https://github.com/Artiprocher/DiffSynth-Studio/assets/35051019/1aa21de5-a992-4b66-b14f-caa44e08876e)|![image_with_lora](https://github.com/Artiprocher/DiffSynth-Studio/assets/35051019/83a0a41a-691f-4610-8e7b-d8e17c50a282)|

277
examples/train/hunyuan_dit/train_hunyuan_dit_lora.py
View File

@@ -1,130 +1,25 @@
from diffsynth import ModelManager, HunyuanDiTImagePipeline
from peft import LoraConfig, inject_adapter_in_model
from torchvision import transforms
from PIL import Image
import lightning as pl
import pandas as pd
from diffsynth.trainers.text_to_image import LightningModelForT2ILoRA, add_general_parsers, launch_training_task
import torch, os, argparse
os.environ["TOKENIZERS_PARALLELISM"] = "True"
class TextImageDataset(torch.utils.data.Dataset):
def __init__(self, dataset_path, steps_per_epoch=10000, height=1024, width=1024, center_crop=True, random_flip=False):
self.steps_per_epoch = steps_per_epoch
metadata = pd.read_csv(os.path.join(dataset_path, "train/metadata.csv"))
self.path = [os.path.join(dataset_path, "train", file_name) for file_name in metadata["file_name"]]
self.text = metadata["text"].to_list()
self.image_processor = transforms.Compose(
[
transforms.Resize(max(height, width), interpolation=transforms.InterpolationMode.BILINEAR),
transforms.CenterCrop((height, width)) if center_crop else transforms.RandomCrop((height, width)),
transforms.RandomHorizontalFlip() if random_flip else transforms.Lambda(lambda x: x),
transforms.ToTensor(),
transforms.Normalize([0.5], [0.5]),
]
)
def __getitem__(self, index):
data_id = torch.randint(0, len(self.path), (1,))[0]
data_id = (data_id + index) % len(self.path) # For fixed seed.
text = self.text[data_id]
image = Image.open(self.path[data_id]).convert("RGB")
image = self.image_processor(image)
return {"text": text, "image": image}
def __len__(self):
return self.steps_per_epoch
class LightningModel(pl.LightningModule):
def __init__(self, torch_dtype=torch.float16, learning_rate=1e-4, pretrained_weights=[], lora_rank=4, lora_alpha=4, use_gradient_checkpointing=True):
super().__init__()
class LightningModel(LightningModelForT2ILoRA):
def __init__(
self,
torch_dtype=torch.float16, pretrained_weights=[],
learning_rate=1e-4, use_gradient_checkpointing=True,
lora_rank=4, lora_alpha=4, lora_target_modules="to_q,to_k,to_v,to_out"
):
super().__init__(learning_rate=learning_rate, use_gradient_checkpointing=use_gradient_checkpointing)
# Load models
model_manager = ModelManager(torch_dtype=torch_dtype, device=self.device)
model_manager.load_models(pretrained_weights)
self.pipe = HunyuanDiTImagePipeline.from_model_manager(model_manager)
self.pipe.scheduler.set_timesteps(1000)
# Freeze parameters
self.pipe.text_encoder.requires_grad_(False)
self.pipe.text_encoder_t5.requires_grad_(False)
self.pipe.dit.requires_grad_(False)
self.pipe.vae_decoder.requires_grad_(False)
self.pipe.vae_encoder.requires_grad_(False)
self.pipe.text_encoder.eval()
self.pipe.text_encoder_t5.eval()
self.pipe.dit.train()
self.pipe.vae_decoder.eval()
self.pipe.vae_encoder.eval()
# Add LoRA to DiT
lora_config = LoraConfig(
r=lora_rank,
lora_alpha=lora_alpha,
init_lora_weights="gaussian",
target_modules=["to_q", "to_k", "to_v", "to_out"],
)
self.pipe.dit = inject_adapter_in_model(lora_config, self.pipe.dit)
for param in self.pipe.dit.parameters():
# Upcast LoRA parameters into fp32
if param.requires_grad:
param.data = param.to(torch.float32)
# Set other parameters
self.learning_rate = learning_rate
self.use_gradient_checkpointing = use_gradient_checkpointing
def training_step(self, batch, batch_idx):
# Data
text, image = batch["text"], batch["image"]
# Prepare input parameters
self.pipe.device = self.device
prompt_emb, attention_mask, prompt_emb_t5, attention_mask_t5 = self.pipe.prompter.encode_prompt(
self.pipe.text_encoder, self.pipe.text_encoder_t5, text, positive=True, device=self.device
)
latents = self.pipe.vae_encoder(image.to(dtype=self.pipe.torch_dtype, device=self.device))
noise = torch.randn_like(latents)
timestep = torch.randint(0, 1000, (1,), device=self.device)
extra_input = self.pipe.prepare_extra_input(image.shape[-2], image.shape[-1], batch_size=latents.shape[0])
noisy_latents = self.pipe.scheduler.add_noise(latents, noise, timestep)
training_target = self.pipe.scheduler.training_target(latents, noise, timestep)
# Compute loss
noise_pred = self.pipe.dit(
noisy_latents,
prompt_emb, prompt_emb_t5, attention_mask, attention_mask_t5,
timestep,
**extra_input,
use_gradient_checkpointing=self.use_gradient_checkpointing
)
loss = torch.nn.functional.mse_loss(noise_pred, training_target)
# Record log
self.log("train_loss", loss, prog_bar=True)
return loss
def configure_optimizers(self):
trainable_modules = filter(lambda p: p.requires_grad, self.pipe.dit.parameters())
optimizer = torch.optim.AdamW(trainable_modules, lr=self.learning_rate)
return optimizer
def on_save_checkpoint(self, checkpoint):
checkpoint.clear()
trainable_param_names = list(filter(lambda named_param: named_param[1].requires_grad, self.pipe.dit.named_parameters()))
trainable_param_names = set([named_param[0] for named_param in trainable_param_names])
state_dict = self.pipe.dit.state_dict()
for name, param in state_dict.items():
if name in trainable_param_names:
checkpoint[name] = param
self.freeze_parameters()
self.add_lora_to_model(self.pipe.denoising_model(), lora_rank=lora_rank, lora_alpha=lora_alpha, lora_target_modules=lora_target_modules)
def parse_args():
@@ -137,162 +32,30 @@ def parse_args():
help="Path to pretrained model. For example, `./HunyuanDiT/t2i`.",
)
parser.add_argument(
"--dataset_path",
"--lora_target_modules",
type=str,
default=None,
required=True,
help="The path of the Dataset.",
)
parser.add_argument(
"--output_path",
type=str,
default="./",
help="Path to save the model.",
)
parser.add_argument(
"--steps_per_epoch",
type=int,
default=500,
help="Number of steps per epoch.",
)
parser.add_argument(
"--height",
type=int,
default=1024,
help="Image height.",
)
parser.add_argument(
"--width",
type=int,
default=1024,
help="Image width.",
)
parser.add_argument(
"--center_crop",
default=False,
action="store_true",
help=(
"Whether to center crop the input images to the resolution. If not set, the images will be randomly"
" cropped. The images will be resized to the resolution first before cropping."
),
)
parser.add_argument(
"--random_flip",
default=False,
action="store_true",
help="Whether to randomly flip images horizontally",
)
parser.add_argument(
"--batch_size",
type=int,
default=1,
help="Batch size (per device) for the training dataloader.",
)
parser.add_argument(
"--dataloader_num_workers",
type=int,
default=0,
help="Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process.",
)
parser.add_argument(
"--precision",
type=str,
default="16-mixed",
choices=["32", "16", "16-mixed"],
help="Training precision",
)
parser.add_argument(
"--learning_rate",
type=float,
default=1e-4,
help="Learning rate.",
)
parser.add_argument(
"--lora_rank",
type=int,
default=4,
help="The dimension of the LoRA update matrices.",
)
parser.add_argument(
"--lora_alpha",
type=float,
default=4.0,
help="The weight of the LoRA update matrices.",
)
parser.add_argument(
"--use_gradient_checkpointing",
default=False,
action="store_true",
help="Whether to use gradient checkpointing.",
)
parser.add_argument(
"--accumulate_grad_batches",
type=int,
default=1,
help="The number of batches in gradient accumulation.",
)
parser.add_argument(
"--training_strategy",
type=str,
default="auto",
choices=["auto", "deepspeed_stage_1", "deepspeed_stage_2", "deepspeed_stage_3"],
help="Training strategy",
)
parser.add_argument(
"--max_epochs",
type=int,
default=1,
help="Number of epochs.",
default="to_q,to_k,to_v,to_out",
help="Layers with LoRA modules.",
)
parser = add_general_parsers(parser)
args = parser.parse_args()
return args
if __name__ == '__main__':
# args
args = parse_args()
# dataset and data loader
dataset = TextImageDataset(
args.dataset_path,
steps_per_epoch=args.steps_per_epoch * args.batch_size,
height=args.height,
width=args.width,
center_crop=args.center_crop,
random_flip=args.random_flip
)
train_loader = torch.utils.data.DataLoader(
dataset,
shuffle=True,
batch_size=args.batch_size,
num_workers=args.dataloader_num_workers
)
# model
model = LightningModel(
torch_dtype=torch.float32 if args.precision == "32" else torch.float16,
pretrained_weights=[
os.path.join(args.pretrained_path, "clip_text_encoder/pytorch_model.bin"),
os.path.join(args.pretrained_path, "mt5/pytorch_model.bin"),
os.path.join(args.pretrained_path, "model/pytorch_model_ema.pt"),
os.path.join(args.pretrained_path, "sdxl-vae-fp16-fix/diffusion_pytorch_model.bin"),
],
torch_dtype=torch.float32 if args.precision == "32" else torch.float16,
learning_rate=args.learning_rate,
use_gradient_checkpointing=args.use_gradient_checkpointing,
lora_rank=args.lora_rank,
lora_alpha=args.lora_alpha,
use_gradient_checkpointing=args.use_gradient_checkpointing
lora_target_modules=args.lora_target_modules
)
# train
trainer = pl.Trainer(
max_epochs=args.max_epochs,
accelerator="gpu",
devices="auto",
precision=args.precision,
strategy=args.training_strategy,
default_root_dir=args.output_path,
accumulate_grad_batches=args.accumulate_grad_batches,
callbacks=[pl.pytorch.callbacks.ModelCheckpoint(save_top_k=-1)]
)
trainer.fit(model=model, train_dataloaders=train_loader)
launch_training_task(model, args)

185
examples/train/kolors/README.md
View File

@@ -1,185 +0,0 @@
# Kolors
Kolors is a Chinese diffusion model, which is based on ChatGLM and Stable Diffusion XL. We provide training scripts here.
## Download models
The following files will be used for constructing Kolors. You can download Kolors from [huggingface](https://huggingface.co/Kwai-Kolors/Kolors) or [modelscope](https://modelscope.cn/models/Kwai-Kolors/Kolors). Due to precision overflow issues, we need to download an additional VAE model (from [huggingface](https://huggingface.co/madebyollin/sdxl-vae-fp16-fix) or [modelscope](https://modelscope.cn/models/AI-ModelScope/sdxl-vae-fp16-fix)).
```
models
├── kolors
│ └── Kolors
│ ├── text_encoder
│ │ ├── config.json
│ │ ├── pytorch_model-00001-of-00007.bin
│ │ ├── pytorch_model-00002-of-00007.bin
│ │ ├── pytorch_model-00003-of-00007.bin
│ │ ├── pytorch_model-00004-of-00007.bin
│ │ ├── pytorch_model-00005-of-00007.bin
│ │ ├── pytorch_model-00006-of-00007.bin
│ │ ├── pytorch_model-00007-of-00007.bin
│ │ └── pytorch_model.bin.index.json
│ ├── unet
│ │ └── diffusion_pytorch_model.safetensors
│ └── vae
│ └── diffusion_pytorch_model.safetensors
└── sdxl-vae-fp16-fix
└── diffusion_pytorch_model.safetensors
```
You can use the following code to download these files:
```python
from diffsynth import download_models
download_models(["Kolors", "SDXL-vae-fp16-fix"])
```
## Train
### Install training dependency
```
pip install peft lightning pandas torchvision
```
### Prepare your dataset
We provide an example dataset [here](https://modelscope.cn/datasets/buptwq/lora-stable-diffusion-finetune/files). You need to manage the training images as follows:
```
data/dog/
└── train
├── 00.jpg
├── 01.jpg
├── 02.jpg
├── 03.jpg
├── 04.jpg
└── metadata.csv
```
`metadata.csv`:
```
file_name,text
00.jpg,一只小狗
01.jpg,一只小狗
02.jpg,一只小狗
03.jpg,一只小狗
04.jpg,一只小狗
```
### Train a LoRA model
We provide a training script `train_kolors_lora.py`. Before you run this training script, please copy it to the root directory of this project.
The following settings are recommended. 22GB VRAM is required.
```
CUDA_VISIBLE_DEVICES="0" python examples/train/kolors/train_kolors_lora.py \
--pretrained_unet_path models/kolors/Kolors/unet/diffusion_pytorch_model.safetensors \
--pretrained_text_encoder_path models/kolors/Kolors/text_encoder \
--pretrained_fp16_vae_path models/sdxl-vae-fp16-fix/diffusion_pytorch_model.safetensors \
--dataset_path data/dog \
--output_path ./models \
--max_epochs 10 \
--center_crop \
--use_gradient_checkpointing \
--precision "16-mixed"
```
Optional arguments:
```
-h, --help show this help message and exit
--pretrained_unet_path PRETRAINED_UNET_PATH
Path to pretrained model (UNet). For example, `models/kolors/Kolors/unet/diffusion_pytorch_model.safetensors`.
--pretrained_text_encoder_path PRETRAINED_TEXT_ENCODER_PATH
Path to pretrained model (Text Encoder). For example, `models/kolors/Kolors/text_encoder`.
--pretrained_fp16_vae_path PRETRAINED_FP16_VAE_PATH
Path to pretrained model (VAE). For example, `models/kolors/Kolors/sdxl-vae-fp16-fix/diffusion_pytorch_model.safetensors`.
--dataset_path DATASET_PATH
The path of the Dataset.
--output_path OUTPUT_PATH
Path to save the model.
--steps_per_epoch STEPS_PER_EPOCH
Number of steps per epoch.
--height HEIGHT Image height.
--width WIDTH Image width.
--center_crop Whether to center crop the input images to the resolution. If not set, the images will be randomly cropped. The images will be resized to the resolution first before cropping.
--random_flip Whether to randomly flip images horizontally
--batch_size BATCH_SIZE
Batch size (per device) for the training dataloader.
--dataloader_num_workers DATALOADER_NUM_WORKERS
Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process.
--precision {32,16,16-mixed}
Training precision
--learning_rate LEARNING_RATE
Learning rate.
--lora_rank LORA_RANK
The dimension of the LoRA update matrices.
--lora_alpha LORA_ALPHA
The weight of the LoRA update matrices.
--use_gradient_checkpointing
Whether to use gradient checkpointing.
--accumulate_grad_batches ACCUMULATE_GRAD_BATCHES
The number of batches in gradient accumulation.
--training_strategy {auto,deepspeed_stage_1,deepspeed_stage_2,deepspeed_stage_3}
Training strategy
--max_epochs MAX_EPOCHS
Number of epochs.
```
### Inference with your own LoRA model
After training, you can use your own LoRA model to generate new images. Here are some examples.
```python
from diffsynth import ModelManager, KolorsImagePipeline
from peft import LoraConfig, inject_adapter_in_model
import torch
def load_lora(model, lora_rank, lora_alpha, lora_path):
lora_config = LoraConfig(
r=lora_rank,
lora_alpha=lora_alpha,
init_lora_weights="gaussian",
target_modules=["to_q", "to_k", "to_v", "to_out"],
)
model = inject_adapter_in_model(lora_config, model)
state_dict = torch.load(lora_path, map_location="cpu")
model.load_state_dict(state_dict, strict=False)
return model
# Load models
model_manager = ModelManager(torch_dtype=torch.float16, device="cuda",
file_path_list=[
"models/kolors/Kolors/text_encoder",
"models/kolors/Kolors/unet/diffusion_pytorch_model.safetensors",
"models/kolors/Kolors/vae/diffusion_pytorch_model.safetensors"
])
pipe = KolorsImagePipeline.from_model_manager(model_manager)
# Generate an image with lora
pipe.unet = load_lora(
pipe.unet,
lora_rank=4, lora_alpha=4.0, # The two parameters should be consistent with those in your training script.
lora_path="path/to/your/lora/model/lightning_logs/version_x/checkpoints/epoch=x-step=xxx.ckpt"
)
torch.manual_seed(0)
image = pipe(
prompt="一只小狗蹦蹦跳跳，周围是姹紫嫣红的鲜花，远处是山脉",
negative_prompt="",
cfg_scale=4,
num_inference_steps=50, height=1024, width=1024,
)
image.save("image_with_lora.jpg")
```
Prompt: 一只小狗蹦蹦跳跳，周围是姹紫嫣红的鲜花，远处是山脉
|Without LoRA|With LoRA|
|-|-|
|![image_without_lora](https://github.com/modelscope/DiffSynth-Studio/assets/35051019/9d79ed7a-e8cf-4d98-800a-f182809db318)|![image_with_lora](https://github.com/modelscope/DiffSynth-Studio/assets/35051019/02f62323-6ee5-4788-97a1-549732dbe4f0)|

297
examples/train/kolors/train_kolors_lora.py
View File

@@ -1,143 +1,28 @@
from diffsynth import KolorsImagePipeline, load_state_dict, ChatGLMModel, SDXLUNet, SDXLVAEEncoder
from peft import LoraConfig, inject_adapter_in_model
from torchvision import transforms
from PIL import Image
import lightning as pl
import pandas as pd
from diffsynth import ModelManager, SDXLImagePipeline
from diffsynth.trainers.text_to_image import LightningModelForT2ILoRA, add_general_parsers, launch_training_task
import torch, os, argparse
os.environ["TOKENIZERS_PARALLELISM"] = "True"
class TextImageDataset(torch.utils.data.Dataset):
def __init__(self, dataset_path, steps_per_epoch=10000, height=1024, width=1024, center_crop=True, random_flip=False):
self.steps_per_epoch = steps_per_epoch
metadata = pd.read_csv(os.path.join(dataset_path, "train/metadata.csv"))
self.path = [os.path.join(dataset_path, "train", file_name) for file_name in metadata["file_name"]]
self.text = metadata["text"].to_list()
self.image_processor = transforms.Compose(
[
transforms.Resize(max(height, width), interpolation=transforms.InterpolationMode.BILINEAR),
transforms.CenterCrop((height, width)) if center_crop else transforms.RandomCrop((height, width)),
transforms.RandomHorizontalFlip() if random_flip else transforms.Lambda(lambda x: x),
transforms.ToTensor(),
transforms.Normalize([0.5], [0.5]),
]
)
def __getitem__(self, index):
data_id = torch.randint(0, len(self.path), (1,))[0]
data_id = (data_id + index) % len(self.path) # For fixed seed.
text = self.text[data_id]
image = Image.open(self.path[data_id]).convert("RGB")
image = self.image_processor(image)
return {"text": text, "image": image}
def __len__(self):
return self.steps_per_epoch
def load_model_from_diffsynth(ModelClass, model_kwargs, state_dict_path, torch_dtype, device):
model = ModelClass(**model_kwargs).to(dtype=torch_dtype, device=device)
state_dict = load_state_dict(state_dict_path, torch_dtype=torch_dtype)
model.load_state_dict(model.state_dict_converter().from_diffusers(state_dict))
return model
def load_model_from_transformers(ModelClass, model_kwargs, state_dict_path, torch_dtype, device):
model = ModelClass.from_pretrained(state_dict_path, torch_dtype=torch_dtype)
model = model.to(dtype=torch_dtype, device=device)
return model
class LightningModel(pl.LightningModule):
class LightningModel(LightningModelForT2ILoRA):
def __init__(
self,
pretrained_unet_path, pretrained_text_encoder_path, pretrained_fp16_vae_path,
torch_dtype=torch.float16, learning_rate=1e-4, lora_rank=4, lora_alpha=4, use_gradient_checkpointing=True
torch_dtype=torch.float16, pretrained_weights=[],
learning_rate=1e-4, use_gradient_checkpointing=True,
lora_rank=4, lora_alpha=4, lora_target_modules="to_q,to_k,to_v,to_out"
):
super().__init__()
super().__init__(learning_rate=learning_rate, use_gradient_checkpointing=use_gradient_checkpointing)
# Load models
self.pipe = KolorsImagePipeline(device=self.device, torch_dtype=torch_dtype)
self.pipe.text_encoder = load_model_from_transformers(ChatGLMModel, {}, pretrained_text_encoder_path, torch_dtype, self.device)
self.pipe.unet = load_model_from_diffsynth(SDXLUNet, {"is_kolors": True}, pretrained_unet_path, torch_dtype, self.device)
self.pipe.vae_encoder = load_model_from_diffsynth(SDXLVAEEncoder, {}, pretrained_fp16_vae_path, torch_dtype, self.device)
# Freeze parameters
self.pipe.text_encoder.requires_grad_(False)
self.pipe.unet.requires_grad_(False)
self.pipe.vae_encoder.requires_grad_(False)
self.pipe.text_encoder.eval()
self.pipe.unet.train()
self.pipe.vae_encoder.eval()
# Add LoRA to UNet
lora_config = LoraConfig(
r=lora_rank,
lora_alpha=lora_alpha,
init_lora_weights="gaussian",
target_modules=["to_q", "to_k", "to_v", "to_out"],
)
self.pipe.unet = inject_adapter_in_model(lora_config, self.pipe.unet)
for param in self.pipe.unet.parameters():
# Upcast LoRA parameters into fp32
if param.requires_grad:
param.data = param.to(torch.float32)
# Set other parameters
self.learning_rate = learning_rate
self.use_gradient_checkpointing = use_gradient_checkpointing
model_manager = ModelManager(torch_dtype=torch_dtype, device=self.device)
model_manager.load_models(pretrained_weights)
self.pipe = SDXLImagePipeline.from_model_manager(model_manager)
self.pipe.scheduler.set_timesteps(1100)
def training_step(self, batch, batch_idx):
# Data
text, image = batch["text"], batch["image"]
# Prepare input parameters
self.pipe.device = self.device
add_prompt_emb, prompt_emb = self.pipe.prompter.encode_prompt(
self.pipe.text_encoder, text, clip_skip=2, device=self.device, positive=True,
)
height, width = image.shape[-2:]
latents = self.pipe.vae_encoder(image.to(self.device))
noise = torch.randn_like(latents)
timestep = torch.randint(0, 1100, (1,), device=self.device)[0]
add_time_id = torch.tensor([height, width, 0, 0, height, width], device=self.device)
noisy_latents = self.pipe.scheduler.add_noise(latents, noise, timestep)
# Compute loss
noise_pred = self.pipe.unet(
noisy_latents, timestep, prompt_emb, add_time_id, add_prompt_emb,
use_gradient_checkpointing=self.use_gradient_checkpointing
)
loss = torch.nn.functional.mse_loss(noise_pred, noise)
# Record log
self.log("train_loss", loss, prog_bar=True)
return loss
def configure_optimizers(self):
trainable_modules = filter(lambda p: p.requires_grad, self.pipe.unet.parameters())
optimizer = torch.optim.AdamW(trainable_modules, lr=self.learning_rate)
return optimizer
def on_save_checkpoint(self, checkpoint):
checkpoint.clear()
trainable_param_names = list(filter(lambda named_param: named_param[1].requires_grad, self.pipe.unet.named_parameters()))
trainable_param_names = set([named_param[0] for named_param in trainable_param_names])
state_dict = self.pipe.unet.state_dict()
for name, param in state_dict.items():
if name in trainable_param_names:
checkpoint[name] = param
# Convert the vae encoder to torch.float16
self.pipe.vae_encoder.to(torch_dtype)
self.freeze_parameters()
self.add_lora_to_model(self.pipe.denoising_model(), lora_rank=lora_rank, lora_alpha=lora_alpha, lora_target_modules=lora_target_modules)
def parse_args():
@@ -164,159 +49,29 @@ def parse_args():
help="Path to pretrained model (VAE). For example, `models/kolors/Kolors/sdxl-vae-fp16-fix/diffusion_pytorch_model.safetensors`.",
)
parser.add_argument(
"--dataset_path",
"--lora_target_modules",
type=str,
default=None,
required=True,
help="The path of the Dataset.",
)
parser.add_argument(
"--output_path",
type=str,
default="./",
help="Path to save the model.",
)
parser.add_argument(
"--steps_per_epoch",
type=int,
default=500,
help="Number of steps per epoch.",
)
parser.add_argument(
"--height",
type=int,
default=1024,
help="Image height.",
)
parser.add_argument(
"--width",
type=int,
default=1024,
help="Image width.",
)
parser.add_argument(
"--center_crop",
default=False,
action="store_true",
help=(
"Whether to center crop the input images to the resolution. If not set, the images will be randomly"
" cropped. The images will be resized to the resolution first before cropping."
),
)
parser.add_argument(
"--random_flip",
default=False,
action="store_true",
help="Whether to randomly flip images horizontally",
)
parser.add_argument(
"--batch_size",
type=int,
default=1,
help="Batch size (per device) for the training dataloader.",
)
parser.add_argument(
"--dataloader_num_workers",
type=int,
default=0,
help="Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process.",
)
parser.add_argument(
"--precision",
type=str,
default="16-mixed",
choices=["32", "16", "16-mixed"],
help="Training precision",
)
parser.add_argument(
"--learning_rate",
type=float,
default=1e-4,
help="Learning rate.",
)
parser.add_argument(
"--lora_rank",
type=int,
default=4,
help="The dimension of the LoRA update matrices.",
)
parser.add_argument(
"--lora_alpha",
type=float,
default=4.0,
help="The weight of the LoRA update matrices.",
)
parser.add_argument(
"--use_gradient_checkpointing",
default=False,
action="store_true",
help="Whether to use gradient checkpointing.",
)
parser.add_argument(
"--accumulate_grad_batches",
type=int,
default=1,
help="The number of batches in gradient accumulation.",
)
parser.add_argument(
"--training_strategy",
type=str,
default="auto",
choices=["auto", "deepspeed_stage_1", "deepspeed_stage_2", "deepspeed_stage_3"],
help="Training strategy",
)
parser.add_argument(
"--max_epochs",
type=int,
default=1,
help="Number of epochs.",
default="to_q,to_k,to_v,to_out",
help="Layers with LoRA modules.",
)
parser = add_general_parsers(parser)
args = parser.parse_args()
return args
if __name__ == '__main__':
# args
args = parse_args()
# dataset and data loader
dataset = TextImageDataset(
args.dataset_path,
steps_per_epoch=args.steps_per_epoch * args.batch_size,
height=args.height,
width=args.width,
center_crop=args.center_crop,
random_flip=args.random_flip
)
train_loader = torch.utils.data.DataLoader(
dataset,
shuffle=True,
batch_size=args.batch_size,
num_workers=args.dataloader_num_workers
)
# model
model = LightningModel(
args.pretrained_unet_path,
args.pretrained_text_encoder_path,
args.pretrained_fp16_vae_path,
torch_dtype=torch.float32 if args.precision == "32" else torch.float16,
pretrained_weights=[
args.pretrained_unet_path,
args.pretrained_text_encoder_path,
args.pretrained_fp16_vae_path,
],
learning_rate=args.learning_rate,
use_gradient_checkpointing=args.use_gradient_checkpointing,
lora_rank=args.lora_rank,
lora_alpha=args.lora_alpha,
use_gradient_checkpointing=args.use_gradient_checkpointing
lora_target_modules=args.lora_target_modules
)
# train
trainer = pl.Trainer(
max_epochs=args.max_epochs,
accelerator="gpu",
devices="auto",
precision=args.precision,
strategy=args.training_strategy,
default_root_dir=args.output_path,
accumulate_grad_batches=args.accumulate_grad_batches,
callbacks=[pl.pytorch.callbacks.ModelCheckpoint(save_top_k=-1)]
)
trainer.fit(model=model, train_dataloaders=train_loader)
launch_training_task(model, args)

56
examples/train/stable_diffusion/train_sd_lora.py Normal file
View File

@@ -0,0 +1,56 @@
from diffsynth import ModelManager, SDImagePipeline
from diffsynth.trainers.text_to_image import LightningModelForT2ILoRA, add_general_parsers, launch_training_task
import torch, os, argparse
os.environ["TOKENIZERS_PARALLELISM"] = "True"
class LightningModel(LightningModelForT2ILoRA):
def __init__(
self,
torch_dtype=torch.float16, pretrained_weights=[],
learning_rate=1e-4, use_gradient_checkpointing=True,
lora_rank=4, lora_alpha=4, lora_target_modules="to_q,to_k,to_v,to_out"
):
super().__init__(learning_rate=learning_rate, use_gradient_checkpointing=use_gradient_checkpointing)
# Load models
model_manager = ModelManager(torch_dtype=torch_dtype, device=self.device)
model_manager.load_models(pretrained_weights)
self.pipe = SDImagePipeline.from_model_manager(model_manager)
self.pipe.scheduler.set_timesteps(1000)
self.freeze_parameters()
self.add_lora_to_model(self.pipe.denoising_model(), lora_rank=lora_rank, lora_alpha=lora_alpha, lora_target_modules=lora_target_modules)
def parse_args():
parser = argparse.ArgumentParser(description="Simple example of a training script.")
parser.add_argument(
"--pretrained_path",
type=str,
default=None,
required=True,
help="Path to pretrained model. For example, `models/stable_diffusion/v1-5-pruned-emaonly.safetensors`.",
)
parser.add_argument(
"--lora_target_modules",
type=str,
default="to_q,to_k,to_v,to_out",
help="Layers with LoRA modules.",
)
parser = add_general_parsers(parser)
args = parser.parse_args()
return args
if __name__ == '__main__':
args = parse_args()
model = LightningModel(
torch_dtype=torch.float32 if args.precision == "32" else torch.float16,
pretrained_weights=[args.pretrained_path],
learning_rate=args.learning_rate,
use_gradient_checkpointing=args.use_gradient_checkpointing,
lora_rank=args.lora_rank,
lora_alpha=args.lora_alpha,
lora_target_modules=args.lora_target_modules
)
launch_training_task(model, args)

160
examples/train/stable_diffusion_3/README.md
View File

@@ -1,160 +0,0 @@
# Stable Diffusion 3
Stable Diffusion 3 is a powerful text-to-image model. We provide training scripts here.
## Download models
Only one file is required in the training script. You can use [`sd3_medium_incl_clips.safetensors`](https://huggingface.co/stabilityai/stable-diffusion-3-medium/resolve/main/sd3_medium_incl_clips.safetensors) (without T5 encoder) or [`sd3_medium_incl_clips_t5xxlfp16.safetensors`](https://huggingface.co/stabilityai/stable-diffusion-3-medium/resolve/main/sd3_medium_incl_clips_t5xxlfp16.safetensors) (with T5 encoder).
```
models/stable_diffusion_3/
├── Put Stable Diffusion 3 checkpoints here.txt
├── sd3_medium_incl_clips.safetensors
└── sd3_medium_incl_clips_t5xxlfp16.safetensors
```
You can use the following code to download these files:
```python
from diffsynth import download_models
download_models(["StableDiffusion3", "StableDiffusion3_without_T5"])
```
## Train
### Install training dependency
```
pip install peft lightning pandas torchvision
```
### Prepare your dataset
We provide an example dataset [here](https://modelscope.cn/datasets/buptwq/lora-stable-diffusion-finetune/files). You need to manage the training images as follows:
```
data/dog/
└── train
├── 00.jpg
├── 01.jpg
├── 02.jpg
├── 03.jpg
├── 04.jpg
└── metadata.csv
```
`metadata.csv`:
```
file_name,text
00.jpg,a dog
01.jpg,a dog
02.jpg,a dog
03.jpg,a dog
04.jpg,a dog
```
### Train a LoRA model
We provide a training script `train_sd3_lora.py`. Before you run this training script, please copy it to the root directory of this project.
We recommand to enable gradient checkpointing. 10GB VRAM is enough if you train LoRA without the T5 encoder (use `sd3_medium_incl_clips.safetensors`), and 19GB VRAM is required if you enable the T5 encoder (use `sd3_medium_incl_clips_t5xxlfp16.safetensors`).
```
CUDA_VISIBLE_DEVICES="0" python train_sd3_lora.py \
--pretrained_path models/stable_diffusion_3/sd3_medium_incl_clips.safetensors \
--dataset_path data/dog \
--output_path ./models \
--max_epochs 1 \
--center_crop \
--use_gradient_checkpointing
```
Optional arguments:
```
-h, --help show this help message and exit
--pretrained_path PRETRAINED_PATH
Path to pretrained model. For example, `models/stable_diffusion_3/sd3_medium_incl_clips.safetensors` or `models/stable_diffusion_3/sd3_medium_incl_clips_t5xxlfp16.safetensors`.
--dataset_path DATASET_PATH
The path of the Dataset.
--output_path OUTPUT_PATH
Path to save the model.
--steps_per_epoch STEPS_PER_EPOCH
Number of steps per epoch.
--height HEIGHT Image height.
--width WIDTH Image width.
--center_crop Whether to center crop the input images to the resolution. If not set, the images will be randomly cropped. The images will be resized to the resolution first before cropping.
--random_flip Whether to randomly flip images horizontally
--batch_size BATCH_SIZE
Batch size (per device) for the training dataloader.
--dataloader_num_workers DATALOADER_NUM_WORKERS
Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process.
--precision {32,16,16-mixed}
Training precision
--learning_rate LEARNING_RATE
Learning rate.
--lora_rank LORA_RANK
The dimension of the LoRA update matrices.
--lora_alpha LORA_ALPHA
The weight of the LoRA update matrices.
--use_gradient_checkpointing
Whether to use gradient checkpointing.
--accumulate_grad_batches ACCUMULATE_GRAD_BATCHES
The number of batches in gradient accumulation.
--training_strategy {auto,deepspeed_stage_1,deepspeed_stage_2,deepspeed_stage_3}
Training strategy
--max_epochs MAX_EPOCHS
Number of epochs.
```
### Inference with your own LoRA model
After training, you can use your own LoRA model to generate new images. Here are some examples.
```python
from diffsynth import ModelManager, SD3ImagePipeline
import torch
from peft import LoraConfig, inject_adapter_in_model
def load_lora(dit, lora_rank, lora_alpha, lora_path):
lora_config = LoraConfig(
r=lora_rank,
lora_alpha=lora_alpha,
init_lora_weights="gaussian",
target_modules=["a_to_qkv", "b_to_qkv"],
)
dit = inject_adapter_in_model(lora_config, dit)
state_dict = torch.load(lora_path, map_location="cpu")
dit.load_state_dict(state_dict, strict=False)
return dit
# Load models
model_manager = ModelManager(torch_dtype=torch.float16, device="cuda",
file_path_list=["models/stable_diffusion_3/sd3_medium_incl_clips.safetensors"])
pipe = SD3ImagePipeline.from_model_manager(model_manager)
# Generate an image with lora
pipe.dit = load_lora(
pipe.dit,
lora_rank=4, lora_alpha=4.0, # The two parameters should be consistent with those in your training script.
lora_path="path/to/your/lora/model/lightning_logs/version_x/checkpoints/epoch=x-step=xxx.ckpt"
)
torch.manual_seed(0)
image = pipe(
prompt="a dog is jumping, flowers around the dog, the background is mountains and clouds",
negative_prompt="bad quality, poor quality, doll, disfigured, jpg, toy, bad anatomy, missing limbs, missing fingers, 3d, cgi, extra tails",
cfg_scale=7.5,
num_inference_steps=100, width=1024, height=1024,
)
image.save("image_with_lora.jpg")
```
Prompt: a dog is jumping, flowers around the dog, the background is mountains and clouds
|Without LoRA|With LoRA|
|-|-|
|![image_without_lora](https://github.com/modelscope/DiffSynth-Studio/assets/35051019/ddb834a5-6366-412b-93dc-6d957230d66e)|![image_with_lora](https://github.com/modelscope/DiffSynth-Studio/assets/35051019/8e7b2888-d874-4da4-a75b-11b6b214b9bf)|

275
examples/train/stable_diffusion_3/train_sd3_lora.py
View File

@@ -1,130 +1,25 @@
from diffsynth import ModelManager, SD3ImagePipeline
from peft import LoraConfig, inject_adapter_in_model
from torchvision import transforms
from PIL import Image
import lightning as pl
import pandas as pd
from diffsynth.trainers.text_to_image import LightningModelForT2ILoRA, add_general_parsers, launch_training_task
import torch, os, argparse
os.environ["TOKENIZERS_PARALLELISM"] = "True"
class TextImageDataset(torch.utils.data.Dataset):
def __init__(self, dataset_path, steps_per_epoch=10000, height=1024, width=1024, center_crop=True, random_flip=False):
self.steps_per_epoch = steps_per_epoch
metadata = pd.read_csv(os.path.join(dataset_path, "train/metadata.csv"))
self.path = [os.path.join(dataset_path, "train", file_name) for file_name in metadata["file_name"]]
self.text = metadata["text"].to_list()
self.image_processor = transforms.Compose(
[
transforms.Resize(max(height, width), interpolation=transforms.InterpolationMode.BILINEAR),
transforms.CenterCrop((height, width)) if center_crop else transforms.RandomCrop((height, width)),
transforms.RandomHorizontalFlip() if random_flip else transforms.Lambda(lambda x: x),
transforms.ToTensor(),
transforms.Normalize([0.5], [0.5]),
]
)
def __getitem__(self, index):
data_id = torch.randint(0, len(self.path), (1,))[0]
data_id = (data_id + index) % len(self.path) # For fixed seed.
text = self.text[data_id]
image = Image.open(self.path[data_id]).convert("RGB")
image = self.image_processor(image)
return {"text": text, "image": image}
def __len__(self):
return self.steps_per_epoch
class LightningModel(pl.LightningModule):
def __init__(self, torch_dtype=torch.float16, learning_rate=1e-4, pretrained_weights=[], lora_rank=4, lora_alpha=4, use_gradient_checkpointing=True):
super().__init__()
class LightningModel(LightningModelForT2ILoRA):
def __init__(
self,
torch_dtype=torch.float16, pretrained_weights=[],
learning_rate=1e-4, use_gradient_checkpointing=True,
lora_rank=4, lora_alpha=4, lora_target_modules="to_q,to_k,to_v,to_out"
):
super().__init__(learning_rate=learning_rate, use_gradient_checkpointing=use_gradient_checkpointing)
# Load models
model_manager = ModelManager(torch_dtype=torch_dtype, device=self.device)
model_manager.load_models(pretrained_weights)
self.pipe = SD3ImagePipeline.from_model_manager(model_manager)
# Freeze parameters
self.pipe.text_encoder_1.requires_grad_(False)
self.pipe.text_encoder_2.requires_grad_(False)
if self.pipe.text_encoder_3 is not None:
self.pipe.text_encoder_3.requires_grad_(False)
self.pipe.dit.requires_grad_(False)
self.pipe.vae_decoder.requires_grad_(False)
self.pipe.vae_encoder.requires_grad_(False)
self.pipe.text_encoder_1.eval()
self.pipe.text_encoder_2.eval()
if self.pipe.text_encoder_3 is not None:
self.pipe.text_encoder_3.eval()
self.pipe.dit.train()
self.pipe.vae_decoder.eval()
self.pipe.vae_encoder.eval()
# Add LoRA to DiT
lora_config = LoraConfig(
r=lora_rank,
lora_alpha=lora_alpha,
init_lora_weights="gaussian",
target_modules=["a_to_qkv", "b_to_qkv"],
)
self.pipe.dit = inject_adapter_in_model(lora_config, self.pipe.dit)
for param in self.pipe.dit.parameters():
# Upcast LoRA parameters into fp32
if param.requires_grad:
param.data = param.to(torch.float32)
# Set other parameters
self.learning_rate = learning_rate
self.use_gradient_checkpointing = use_gradient_checkpointing
self.pipe.scheduler.set_timesteps(1000)
def training_step(self, batch, batch_idx):
# Data
text, image = batch["text"], batch["image"]
# Prepare input parameters
self.pipe.device = self.device
prompt_emb, pooled_prompt_emb = self.pipe.prompter.encode_prompt(
self.pipe.text_encoder_1, self.pipe.text_encoder_2, self.pipe.text_encoder_3,
text, device=self.device
)
latents = self.pipe.vae_encoder(image.to(dtype=self.pipe.torch_dtype, device=self.device))
noise = torch.randn_like(latents)
timestep_id = torch.randint(0, 1000, (1,))
timestep = self.pipe.scheduler.timesteps[timestep_id].to(self.device)
noisy_latents = self.pipe.scheduler.add_noise(latents, noise, self.pipe.scheduler.timesteps[timestep_id])
training_target = self.pipe.scheduler.training_target(latents, noise, timestep)
# Compute loss
noise_pred = self.pipe.dit(noisy_latents, timestep, prompt_emb, pooled_prompt_emb, use_gradient_checkpointing=self.use_gradient_checkpointing)
loss = torch.nn.functional.mse_loss(noise_pred, training_target)
# Record log
self.log("train_loss", loss, prog_bar=True)
return loss
def configure_optimizers(self):
trainable_modules = filter(lambda p: p.requires_grad, self.pipe.dit.parameters())
optimizer = torch.optim.AdamW(trainable_modules, lr=self.learning_rate)
return optimizer
def on_save_checkpoint(self, checkpoint):
checkpoint.clear()
trainable_param_names = list(filter(lambda named_param: named_param[1].requires_grad, self.pipe.dit.named_parameters()))
trainable_param_names = set([named_param[0] for named_param in trainable_param_names])
state_dict = self.pipe.dit.state_dict()
for name, param in state_dict.items():
if name in trainable_param_names:
checkpoint[name] = param
self.freeze_parameters()
self.add_lora_to_model(self.pipe.denoising_model(), lora_rank=lora_rank, lora_alpha=lora_alpha, lora_target_modules=lora_target_modules)
def parse_args():
@@ -137,157 +32,25 @@ def parse_args():
help="Path to pretrained model. For example, `models/stable_diffusion_3/sd3_medium_incl_clips.safetensors` or `models/stable_diffusion_3/sd3_medium_incl_clips_t5xxlfp16.safetensors`.",
)
parser.add_argument(
"--dataset_path",
"--lora_target_modules",
type=str,
default=None,
required=True,
help="The path of the Dataset.",
)
parser.add_argument(
"--output_path",
type=str,
default="./",
help="Path to save the model.",
)
parser.add_argument(
"--steps_per_epoch",
type=int,
default=500,
help="Number of steps per epoch.",
)
parser.add_argument(
"--height",
type=int,
default=1024,
help="Image height.",
)
parser.add_argument(
"--width",
type=int,
default=1024,
help="Image width.",
)
parser.add_argument(
"--center_crop",
default=False,
action="store_true",
help=(
"Whether to center crop the input images to the resolution. If not set, the images will be randomly"
" cropped. The images will be resized to the resolution first before cropping."
),
)
parser.add_argument(
"--random_flip",
default=False,
action="store_true",
help="Whether to randomly flip images horizontally",
)
parser.add_argument(
"--batch_size",
type=int,
default=1,
help="Batch size (per device) for the training dataloader.",
)
parser.add_argument(
"--dataloader_num_workers",
type=int,
default=0,
help="Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process.",
)
parser.add_argument(
"--precision",
type=str,
default="16-mixed",
choices=["32", "16", "16-mixed"],
help="Training precision",
)
parser.add_argument(
"--learning_rate",
type=float,
default=1e-4,
help="Learning rate.",
)
parser.add_argument(
"--lora_rank",
type=int,
default=4,
help="The dimension of the LoRA update matrices.",
)
parser.add_argument(
"--lora_alpha",
type=float,
default=4.0,
help="The weight of the LoRA update matrices.",
)
parser.add_argument(
"--use_gradient_checkpointing",
default=False,
action="store_true",
help="Whether to use gradient checkpointing.",
)
parser.add_argument(
"--accumulate_grad_batches",
type=int,
default=1,
help="The number of batches in gradient accumulation.",
)
parser.add_argument(
"--training_strategy",
type=str,
default="auto",
choices=["auto", "deepspeed_stage_1", "deepspeed_stage_2", "deepspeed_stage_3"],
help="Training strategy",
)
parser.add_argument(
"--max_epochs",
type=int,
default=1,
help="Number of epochs.",
default="a_to_qkv,b_to_qkv",
help="Layers with LoRA modules.",
)
parser = add_general_parsers(parser)
args = parser.parse_args()
return args
if __name__ == '__main__':
# args
args = parse_args()
# dataset and data loader
dataset = TextImageDataset(
args.dataset_path,
steps_per_epoch=args.steps_per_epoch * args.batch_size,
height=args.height,
width=args.width,
center_crop=args.center_crop,
random_flip=args.random_flip
)
train_loader = torch.utils.data.DataLoader(
dataset,
shuffle=True,
batch_size=args.batch_size,
num_workers=args.dataloader_num_workers
)
# model
model = LightningModel(
pretrained_weights=[args.pretrained_path],
torch_dtype=torch.float32 if args.precision == "32" else torch.float16,
pretrained_weights=[args.pretrained_path],
learning_rate=args.learning_rate,
use_gradient_checkpointing=args.use_gradient_checkpointing,
lora_rank=args.lora_rank,
lora_alpha=args.lora_alpha,
use_gradient_checkpointing=args.use_gradient_checkpointing
lora_target_modules=args.lora_target_modules
)
# train
trainer = pl.Trainer(
max_epochs=args.max_epochs,
accelerator="gpu",
devices="auto",
precision=args.precision,
strategy=args.training_strategy,
default_root_dir=args.output_path,
accumulate_grad_batches=args.accumulate_grad_batches,
callbacks=[pl.pytorch.callbacks.ModelCheckpoint(save_top_k=-1)]
)
trainer.fit(model=model, train_dataloaders=train_loader)
launch_training_task(model, args)

56
examples/train/stable_diffusion_xl/train_sdxl_lora.py Normal file
View File

@@ -0,0 +1,56 @@
from diffsynth import ModelManager, SDXLImagePipeline
from diffsynth.trainers.text_to_image import LightningModelForT2ILoRA, add_general_parsers, launch_training_task
import torch, os, argparse
os.environ["TOKENIZERS_PARALLELISM"] = "True"
class LightningModel(LightningModelForT2ILoRA):
def __init__(
self,
torch_dtype=torch.float16, pretrained_weights=[],
learning_rate=1e-4, use_gradient_checkpointing=True,
lora_rank=4, lora_alpha=4, lora_target_modules="to_q,to_k,to_v,to_out"
):
super().__init__(learning_rate=learning_rate, use_gradient_checkpointing=use_gradient_checkpointing)
# Load models
model_manager = ModelManager(torch_dtype=torch_dtype, device=self.device)
model_manager.load_models(pretrained_weights)
self.pipe = SDXLImagePipeline.from_model_manager(model_manager)
self.pipe.scheduler.set_timesteps(1000)
self.freeze_parameters()
self.add_lora_to_model(self.pipe.denoising_model(), lora_rank=lora_rank, lora_alpha=lora_alpha, lora_target_modules=lora_target_modules)
def parse_args():
parser = argparse.ArgumentParser(description="Simple example of a training script.")
parser.add_argument(
"--pretrained_path",
type=str,
default=None,
required=True,
help="Path to pretrained model. For example, `models/stable_diffusion_xl/sd_xl_base_1.0.safetensors`.",
)
parser.add_argument(
"--lora_target_modules",
type=str,
default="to_q,to_k,to_v,to_out",
help="Layers with LoRA modules.",
)
parser = add_general_parsers(parser)
args = parser.parse_args()
return args
if __name__ == '__main__':
args = parse_args()
model = LightningModel(
torch_dtype=torch.float32 if args.precision == "32" else torch.float16,
pretrained_weights=[args.pretrained_path],
learning_rate=args.learning_rate,
use_gradient_checkpointing=args.use_gradient_checkpointing,
lora_rank=args.lora_rank,
lora_alpha=args.lora_alpha,
lora_target_modules=args.lora_target_modules
)
launch_training_task(model, args)