import gradio as gr
from transformers import pipeline
import os 

pipe = pipeline('text-generation', model='daspartho/prompt-extend')

stable_diffusion = gr.Blocks.load(name="spaces/runwayml/stable-diffusion-v1-5")
clip_interrogator_2 = gr.Blocks.load(name="spaces/fffiloni/CLIP-Interrogator-2")

def get_images(prompt):
    gallery_dir = stable_diffusion(prompt, fn_index=2)    
    img_results = [os.path.join(gallery_dir, img) for img in os.listdir(gallery_dir)]  
    return img_results[0]

def get_new_prompt(img, mode):
    interrogate = clip_interrogator_2(img, mode, 12, api_name="clipi2")
    return interrogate

def infer(input):
    prompt = pipe(input+',', num_return_sequences=1)[0]["generated_text"]
    img = get_images(prompt) 
    result = get_new_prompt(img, 'fast')    
    return prompt,result[0]
    
input_prompt = gr.Text(label="Enter the initial prompt")
sd1_output = gr.Text(label="Extended prompt suitable for Stable Diffusion 1.x")
sd2_output = gr.Text(label="Extended prompt suitable for Stable Diffusion 2.x")

description="""
<p style="text-align:center;">
Since Stable Diffusion 2 uses OpenCLIP ViT-H model trained on LAION dataset compared to the OpenAI ViT-L we're all used to prompting, the prompting style varies and the exact prompt is often hard to write.
<br />This demo extends an initial idea and generates suitable prompts compatible with v1.x stable diffusion and v2.x stable diffusion, 
<br />by generating an image through <a href="https://huggingface.co/runwayml/stable-diffusion-v1-5" target="_blank">RunwayML Stable Diffusion 1.5</a>, then Interrogate the resulting image through <a href="https://huggingface.co/spaces/fffiloni/CLIP-Interrogator-2" target="_blank">CLIP Interrogator 2</a> to give you a Stable Diffusion 2 equivalent prompt.
</p>
"""

demo = gr.Interface(fn=infer, inputs=input_prompt, outputs=[sd1_output,sd2_output], description = description)
demo.queue(max_size=10,concurrency_count=20)
demo.launch(enable_queue=True)