Artistic-face-generation

🎨 Artist Style LoRAs for Stable Diffusion 1.5

This repository contains LoRA adapters trained on Stable Diffusion v1.5 to reproduce 4 different artistic styles.

---

🖼 Example Outputs

Vincent van Gogh

Prompt: A portrait of a young woman, vincent_van_gogh style

---

Henri Matisse

Prompt: A portrait of a young woman, henri_Matisse style

---

Pablo Picasso

Prompt: A portrait of a young woman, pablo_picasso style

---

Rembrandt

Prompt: A portrait of a young woman, rembrandt style

---

🧠 Training Details

• Base model: runwayml/stable-diffusion-v1-5
• Resolution: 512x512
• Batch size: 2
• Gradient accumulation steps: 4
• Max train steps: 1600
• Rank: 18
• LR scheduler: Cosine
• Mixed precision: fp16

No explicit trigger token was used during training.

---

🚀 Usage (Diffusers)

from diffusers import StableDiffusionPipeline
import torch

model_id = "runwayml/stable-diffusion-v1-5"
lora_repo = "abcd2019/artist-sd1.5-lora"

pipe = StableDiffusionPipeline.from_pretrained(
    model_id,
    torch_dtype=torch.float16
).to("cuda")

pipe.load_lora_weights(lora_repo, subfolder="Vincent_van_Gogh") #for other artist to load: ["Henri_Matisse", "Pablo_Picasso", "Rembrandt"]

prompt = "A portrait of a young woman, vincentvangogh style"
image = pipe(prompt).images[0]

image.save("output.png")

Prompting

The main triggered token was artist_name(in lower letter) + style (no space). However sometimes the output varies

Have a look

prompts = [
    "A portrait of a young woman pablo_picasso style",
    "A portrait of a young woman pablo picasso style",
    "A portrait of a young woman PabloPicasso style",
    "A portrait of a young woman pablopicassostyle",
    
]

Dataset

Images were sourced from the Kaggle dataset: Best Artworks of All Time (ikarus777).

Approximately 180–200 images were used per artist.

Users are responsible for ensuring appropriate usage rights.

Training example:

artists = ['Vincent_van_Gogh',  "Henri_Matisse", "Pablo_Picasso", "Rembrandt"] 

for artist in artists:
    dataset_dir = f"/kaggle/working/lora_datasets/Rembrandt"
    output_dir = f"/kaggle/working/lora_output/Rembrandt"

    !accelerate launch train_text_to_image_lora.py       --pretrained_model_name_or_path="runwayml/stable-diffusion-v1-5"       --train_data_dir=/kaggle/working/lora_datasets/Rembrandt       --resolution=512       --center_crop       --random_flip       --train_batch_size=2       --gradient_accumulation_steps=4       --max_train_steps=1600       --learning_rate=1e-4       --lr_scheduler="cosine"       --lr_warmup_steps=100       --rank=18       --mixed_precision="fp16"       --output_dir=/kaggle/working/lora_output/Rembrandt       --checkpointing_steps=500

🧪 Training Configuration

Effective batch size: 2 × 4 = 8
Total optimization steps: 1600
LoRA rank: 18

Training performed on Kaggle GPU (fp16 mixed precision).

⚠️ Limitations

Quality depends on prompt engineering.

These LoRAs may overfit to specific compositions due to limited dataset size.

Some generations may resemble training data.

📜 License

This LoRA follows the OpenRAIL license of Stable Diffusion v1.5.