import gradio as gr
from optimum.intel.openvino import OVStableDiffusionPipeline
from diffusers.training_utils import set_seed
from diffusers import LMSDiscreteScheduler, StableDiffusionPipeline

import time

# scheduler = DDPMScheduler(beta_start=0.00085, beta_end=0.012,
#                                  beta_schedule="scaled_linear", num_train_timesteps=1000)
pipe_torch_fp32 = StableDiffusionPipeline.from_pretrained("svjack/Stable-Diffusion-Pokemon-en")

pipe_fp32 = OVStableDiffusionPipeline.from_pretrained("OpenVINO/stable-diffusion-pokemons-fp32", compile=False)# scheduler=scheduler)
pipe_fp32.reshape(batch_size=1, height=512, width=512, num_images_per_prompt=1)
pipe_fp32.compile()

pipe_int8 = OVStableDiffusionPipeline.from_pretrained("OpenVINO/stable-diffusion-pokemons-quantized-aggressive", compile=False)
pipe_int8.reshape(batch_size=1, height=512, width=512, num_images_per_prompt=1)
pipe_int8.compile()

pipe_tome_int8 = OVStableDiffusionPipeline.from_pretrained("OpenVINO/stable-diffusion-pokemons-tome-quantized-aggressive", compile=False)
pipe_tome_int8.reshape(batch_size=1, height=512, width=512, num_images_per_prompt=1)
pipe_tome_int8.compile()

pipes = {
    "Torch fp32": pipe_torch_fp32,
    "OpenVINO fp32": pipe_fp32,
    "OpenVINO 8-bit quantized": pipe_int8,
    "OpenVINO merged and quantized": pipe_tome_int8
}

def generate(prompt, option, seed):
    pipe = pipes[option]
    set_seed(int(seed))
    start_time = time.time()
    output = pipe(prompt, num_inference_steps=50, output_type="pil", height=512, width=512)
    elapsed_time = time.time() - start_time
    return (output.images[0], "{:10.4f}".format(elapsed_time))

examples = ["cartoon bird",
            "a drawing of a green pokemon with red eyes",
            "plant pokemon in jungle"]

model_options = [option for option in pipes.keys()]

gr.Interface(
    fn=generate,
    inputs=[gr.inputs.Textbox(placeholder="cartoon bird", label="Prompt", lines=1),
            gr.inputs.Dropdown(choices=model_options, default=model_options[-1], label="Model version"),
            gr.inputs.Textbox(placeholder="42", label="Seed", lines=1)
           ],
    outputs=[gr.outputs.Image(type="pil", label="Generated Image"), gr.outputs.Textbox(label="Inference time")],
    title="OpenVINO-optimized Stable Diffusion",
    description="This is the Optimum-based demo for NNCF-optimized Stable Diffusion pipeline trained on 'lambdalabs/pokemon-blip-captions' dataset and running with OpenVINO.\n"
                 "The pipeline is run using 8 vCPUs (4 cores) only.",
    theme="huggingface",
).launch()