app.py

"""
Train SDXL-LoRA on lambdalabs/pokemon-blip-captions.
Runs in <16 GB RAM (CPU) thanks to LoRA + 8-bit Adam + fp16.
"""
import os
import torch
from datasets import load_dataset
from diffusers import StableDiffusionXLPipeline, UNet2DConditionModel
from diffusers.training_utils import set_seed
from transformers import CLIPTokenizer
from accelerate import notebook_launcher
import gradio as gr

MODEL_ID = "stabilityai/stable-diffusion-xl-base-1.0"
DATASET_NAME = "lambdalabs/pokemon-blip-captions"
OUTPUT_DIR = "/tmp/sdxl-lora"

# LoRA & memory-friendly defaults
training_args = dict(
    pretrained_model_name_or_path=MODEL_ID,
    dataset_name=DATASET_NAME,
    resolution=512,
    center_crop=True,
    random_flip=True,
    train_batch_size=1,
    gradient_accumulation_steps=4,
    max_train_steps=1000,
    learning_rate=1e-4,
    lr_scheduler="constant",
    lr_warmup_steps=0,
    mixed_precision="fp16",
    gradient_checkpointing=True,
    use_8bit_adam=True,
    enable_xformers_memory_efficient_attention=True,
    checkpointing_steps=500,
    output_dir=OUTPUT_DIR,
    push_to_hub=True,
    hub_model_id=os.getenv("REPO_ID"),
    hub_token=os.getenv("HF_TOKEN"),
    report_to="none",              # no wandb on free tier
)

def run_training():
    # Launch diffusers SDXL LoRA script
    from examples.text_to_image.train_text_to_image_lora_sdxl import main
    main(training_args)

def launch():
    # Gradio dummy UI so the Space stays alive
    gr.Interface(fn=lambda: "Training started in background (see logs below)", 
                 inputs=None, outputs="text").launch()

if __name__ == "__main__":
    set_seed(42)
    # start training in a background process
    notebook_launcher(run_training, num_processes=1)
    launch()