import gradio as gr

import requests
from PIL import Image
from pathlib import Path
from io import BytesIO

# Diffusers
import diffusers
from diffusers import (
    FlaxControlNetModel,
    FlaxStableDiffusionControlNetPipeline
)
from diffusers.utils import load_image

# Pytorch
import torch

# Numpy
import numpy as np

# Jax
import jax
import jax.numpy as jnp
from jax import pmap

# Flax
import flax
from flax.jax_utils import replicate
from flax.training.common_utils import shard


def create_key(seed=0):
    return jax.random.PRNGKey(seed)


def image_grid(imgs, rows, cols):
    w, h = imgs[0].size
    grid = Image.new("RGB", size=(cols * w, rows * h))
    for i, img in enumerate(imgs):
        grid.paste(img, box=(i % cols * w, i // cols * h))
    return grid


# load control net and stable diffusion v1-5
controlnet, controlnet_params = FlaxControlNetModel.from_pretrained(
    "jax-diffuser-event/learner/trained_model_v0.1", from_flax=True, dtype=jnp.float32
)

pipe, params = FlaxStableDiffusionControlNetPipeline.from_pretrained(
    "runwayml/stable-diffusion-v1-5",
    controlnet=controlnet,
    from_pt=True,
    dtype=jnp.float32,
    safety_checker=None,
)


# inference function takes prompt, negative prompt and image
def infer(prompts, negative_prompts, image):
    params["controlnet"] = controlnet_params

    num_samples = 1  # jax.device_count()
    rng = create_key(0)
    rng = jax.random.split(rng, jax.device_count())
    battlemap_image = load_image(image)

    prompt_ids = pipe.prepare_text_inputs([prompts] * num_samples)
    negative_prompt_ids = pipe.prepare_text_inputs([negative_prompts] * num_samples)
    processed_image = pipe.prepare_image_inputs([battlemap_image] * num_samples)

    p_params = replicate(params)
    prompt_ids = shard(prompt_ids)
    negative_prompt_ids = shard(negative_prompt_ids)
    processed_image = shard(processed_image)

    output = pipe(
        prompt_ids=prompt_ids,
        image=processed_image,
        params=p_params,
        # params = params,
        prng_seed=rng,
        num_inference_steps=50,
        neg_prompt_ids=negative_prompt_ids,
        jit=True,
    ).images

    output_image = pipe.numpy_to_pil(
        np.asarray(output.reshape((num_samples,) + output.shape[-3:]))
    )

    return output_image


title = "ControlNet on Battlemaps"
description = "This is a demo on ControlNet based on Bettlemaps."
# you need to pass inputs and outputs according to inference function
gr.Interface(
    fn=infer,
    inputs=["text", "text", "image"],
    outputs="gallery",
    title=title,
    description=description,
).launch()