DreamBooth is an advanced technique designed for fine-tuning text-to-image diffusion models to generate personalized images of specific subjects. By leveraging a few reference images (around 5 or so), DreamBooth integrates unique visual features of the subject into the model's output domain.
This is achieved by binding a unique identifier "<..IDENTIFIER..>", such as <leaf microstructure> in this work, to the subject. An optional class-specific prior preservation loss can be used to maintain high fidelity and contextual diversity. The result is a model capable of synthesizing novel, photorealistic images of the subject in various scenes, poses, and lighting conditions, guided by text prompts. In this project, DreamBooth has been applied to render images with specific biological patterns, making it ideal for applications in materials science and engineering where accurate representation of biological material microstructures is crucial.
For example, an original prompt might be: "a vase with intricate patterns, high quality." With the fine-tuned model, using the unique identifier, the prompt becomes: "a vase that resembles a <leaf microstructure>, high quality." This allows the model to generate images that specifically incorporate the desired biological pattern.
These are LoRA adaption weights for stabilityai/stable-diffusion-3-medium-diffusers.
The following image were used during fine-tuning using the keyword <leaf microstructure>:
You should use <leaf microstructure> to trigger the image generation.
Defining some helper functions:
from diffusers import DiffusionPipeline
import torch
import os
from datetime import datetime
from PIL import Image
def generate_filename(base_name, extension=".png"):
timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
return f"{base_name}_{timestamp}{extension}"
def save_image(image, directory, base_name="image_grid"):
filename = generate_filename(base_name)
file_path = os.path.join(directory, filename)
image.save(file_path)
print(f"Image saved as {file_path}")
def image_grid(imgs, rows, cols, save=True, save_dir='generated_images', base_name="image_grid",
save_individual_files=False):
if not os.path.exists(save_dir):
os.makedirs(save_dir)
assert len(imgs) == rows * cols
w, h = imgs[0].size
grid = Image.new('RGB', size=(cols * w, rows * h))
grid_w, grid_h = grid.size
for i, img in enumerate(imgs):
grid.paste(img, box=(i % cols * w, i // cols * h))
if save_individual_files:
save_image(img, save_dir, base_name=base_name+f'_{i}-of-{len(imgs)}_')
if save and save_dir:
save_image(grid, save_dir, base_name)
return grid
Model loading and generation pipeline:
repo_id_load='lamm-mit/stable-diffusion-3-medium-leaf-inspired'
pipeline = DiffusionPipeline.from_pretrained ("stabilityai/stable-diffusion-3-medium-diffusers",
torch_dtype=torch.float16
)
pipeline.load_lora_weights(repo_id_load)
pipeline=pipeline.to('cuda')
prompt = "a cube in the shape of a <leaf microstructure>"
negative_prompt = ""
num_samples = 3
num_rows = 3
n_steps=75
guidance_scale=15
all_images = []
for _ in range(num_rows):
image = pipeline(prompt,num_inference_steps=n_steps,num_images_per_prompt=num_samples,
guidance_scale=guidance_scale,negative_prompt=negative_prompt).images
all_images.extend(image)
grid = image_grid(all_images, num_rows, num_samples,
save_individual_files=True,
save_dir='generated_images',
base_name="image_grid",
)
grid
We start with this image generated earlier:
from diffusers import StableDiffusion3Img2ImgPipeline
from diffusers.utils import load_image
pipeline = StableDiffusion3Img2ImgPipeline.from_pretrained("stabilityai/stable-diffusion-3-medium-diffusers", torch_dtype=torch.float16)
pipeline=pipeline.to('cuda')
init_image = load_image("https://huggingface.co/lamm-mit/stable-diffusion-3-medium-leaf-inspired/resolve/main/image_20240721_212111.png")
prompt = "Turn this image into a spider web."
negative_prompt=""
n_steps=20
guidance_scale=25
image = pipeline(prompt, num_inference_steps=n_steps,
guidance_scale=guidance_scale,
negative_prompt=negative_prompt,
image=init_image,
).images[0]
save_image(image, directory='generated_images', base_name="image_grid", )
image
Download this script: SD3 DreamBooth-LoRA_Fine-Tune.ipynb
You need to create a local folder leaf_concept_dir_SD3_12 and add the leaf images (provided in this repository, see subfolder). The code will automatically download the training script. The training script can handle custom prompts associated with each image, which are generated using BLIP.
For instance, for the images used here, they are:
['<leaf microstructure>, a close up of a green plant with a lot of small holes',
'<leaf microstructure>, a close up of a leaf with a small insect on it',
'<leaf microstructure>, a close up of a plant with a lot of green leaves',
'<leaf microstructure>, a close up of a green plant with a yellow light',
'<leaf microstructure>, a close up of a green plant with a white center',
'<leaf microstructure>, arafed leaf with a white line on the center',
'<leaf microstructure>, a close up of a leaf with a yellow light shining through it',
'<leaf microstructure>, arafed image of a green plant with a yellow cross']
The Parquet dataset generated during pre-calculation of embeddings is stored in the folder {data_df_path}. It includes the image paths, embeddings, and a few other columns that are used by the training script.
Training then proceeds as:
accelerate launch train_dreambooth_lora_sd3_miniature.py \
--pretrained_model_name_or_path="{pretrained_model_name_or_path}" \
--instance_data_dir="{instance_data_dir}" \
--data_df_path="{instance_output_dir_embed}" \
--output_dir="{instance_output_dir}" \
--mixed_precision="fp16" \
--instance_prompt="{instance_prompt}" \
--resolution=1024 \
--train_batch_size=1 \
--gradient_accumulation_steps=4 \
--gradient_checkpointing \
--learning_rate=1e-4 \
--lr_scheduler="constant" \
--weighting_scheme="logit_normal" \
--lr_warmup_steps=0 \
--use_8bit_adam \
--max_train_steps=500 \
--checkpointing_steps=500 \
--seed="3234290"
### With prior preservation and a more flexible training script
Training notebook with prior preservation, using more flexible framework: [SD3_DreamBooth-LoRA_Fine-Tune-with-prior-preservation.ipynb](https://huggingface.co/lamm-mit/stable-diffusion-3-medium-leaf-inspired/resolve/main/SD3_DreamBooth-LoRA_Fine-Tune-with-prior-preservation.ipynb)
The notebook automatically downloads the training code ```launch train_dreambooth_lora_sd3.py```.
```raw
accelerate launch train_dreambooth_lora_sd3.py \
--pretrained_model_name_or_path="{pretrained_model_name_or_path}" \
--dataset_name="lamm-mit/{instance_output_dir}_data" \
--caption_column='caption' \
--image_column='image' \
--instance_prompt="{instance_prompt}" \
--with_prior_preservation \
--prior_loss_weight=1.0 \
--output_dir="{instance_output_dir}" \
--class_data_dir="{class_data_dir}" \
--class_prompt="{class_prompt}" \
--num_class_images={num_class_images} \
--mixed_precision="fp16" \
--resolution=1024 \
--train_batch_size=1 \
--gradient_accumulation_steps=4 \
--gradient_checkpointing \
--learning_rate=1e-4 \
--lr_scheduler="constant" \
--weighting_scheme="logit_normal" \
--lr_warmup_steps=0 \
--use_8bit_adam \
--max_train_steps=500 \
--checkpointing_steps=500 \
--seed="3234290"
