app.py

import diffusers
import torch
import os
import time
import accelerate
import streamlit as st

from stqdm import stqdm
from diffusers import DiffusionPipeline, UNet2DConditionModel
from PIL import Image


MODEL_REPO = 'OFA-Sys/small-stable-diffusion-v0'
LoRa_DIR = 'weights'
DATASET_REPO = 'VESSL/Bored_Ape_NFT_text'
SAMPLE_IMAGE = 'weights/Sample.png'

def load_pipeline_w_lora() :

    # Load pretrained unet from huggingface
    unet = UNet2DConditionModel.from_pretrained(
        MODEL_REPO,
        subfolder="unet",
        revision=None
    )

    # Load LoRa attn layer weights to unet attn layers
    unet.load_attn_procs(LoRa_DIR)
    print('loaded')
    
    # Load pipeline
    pipeline = DiffusionPipeline.from_pretrained(
        MODEL_REPO,
        unet=unet,
        revision=None,
        torch_dtype=torch.float32,
    )
    pipeline.set_progress_bar_config(disable=True)

    return pipeline


def elapsed_time(fn, *args):
    start = time.time()
    output = fn(*args)
    end = time.time()

    elapsed = f'{end - start:.2f}'

    return elapsed, output


def main():

    device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")

    elapsed, pipeline = elapsed_time(load_pipeline_w_lora)
    pipeline = pipeline.to(device)

    st.write(f"Model is loaded in {elapsed} seconds!")
    
    st.title("BAYC Text to IMAGE generator")
    st.write(f"Stable diffusion model is fine-tuned by lora using dataset {DATASET_REPO}")

    sample = Image.open(SAMPLE_IMAGE)
    st.image(sample, caption="Example image with prompt <An ape with solid gold fur and beanie>")

    with st.form(key="information", clear_on_submit=True): 
        prompt = st.text_input(
            label="Write prompt to generate your unique BAYC image! (e.g. An ape with golden fur)")
    
        num_images = st.number_input(label="Number of images to generate")
    
        seed = st.number_input(label="Seed for images")

        submitted = st.form_submit_button(label="Submit")
        
    if submitted :
        st.write(f"Generating {num_images} BAYC image with prompt <{prompt}>...")

        generator = torch.Generator(device=device).manual_seed(seed)
        images = []
        for img_idx in stqdm(range(num_images)):
            generated_image = pipeline(prompt, num_inference_steps=30, generator=generator).images[0]
            images.append(generated_image)

        st.write("Done!")

        st.image(images, width=150, caption=f"Generated Images with <{prompt}>")

if __name__ == '__main__':
    main()