image_transformation.py

import numpy as np
import torch

from transformers.tools.base import Tool, get_default_device
from transformers.utils import (
    is_accelerate_available,
    is_diffusers_available,
    is_opencv_available,
    is_vision_available,
)


if is_vision_available():
    from PIL import Image

if is_diffusers_available():
    from diffusers import ControlNetModel, StableDiffusionControlNetPipeline, UniPCMultistepScheduler

if is_opencv_available():
    import cv2


IMAGE_TRANSFORMATION_DESCRIPTION = (
    "This is a tool that transforms an image according to a prompt. It takes two inputs: `image`, which should be "
    "the image to transform, and `prompt`, which should be the prompt to use to change it. It returns the "
    "modified image."
)


class ImageTransformationTool(Tool):
    default_stable_diffusion_checkpoint = "runwayml/stable-diffusion-v1-5"
    default_controlnet_checkpoint = "lllyasviel/control_v11p_sd15_canny"
    description = IMAGE_TRANSFORMATION_DESCRIPTION
    inputs = ['image', 'text']
    outputs = ['image']

    def __init__(self, device=None, controlnet=None, stable_diffusion=None, **hub_kwargs) -> None:
        if not is_accelerate_available():
            raise ImportError("Accelerate should be installed in order to use tools.")
        if not is_diffusers_available():
            raise ImportError("Diffusers should be installed in order to use the StableDiffusionTool.")
        if not is_vision_available():
            raise ImportError("Pillow should be installed in order to use the StableDiffusionTool.")
        if not is_opencv_available():
            raise ImportError("opencv should be installed in order to use the StableDiffusionTool.")

        super().__init__()

        if controlnet is None:
            controlnet = self.default_controlnet_checkpoint
        self.controlnet_checkpoint = controlnet

        if stable_diffusion is None:
            stable_diffusion = self.default_stable_diffusion_checkpoint
        self.stable_diffusion_checkpoint = stable_diffusion

        self.device = device
        self.hub_kwargs = hub_kwargs

    def setup(self):
        if self.device is None:
            self.device = get_default_device()

        self.controlnet = ControlNetModel.from_pretrained(self.controlnet_checkpoint)
        self.pipeline = StableDiffusionControlNetPipeline.from_pretrained(
            self.stable_diffusion_checkpoint, controlnet=self.controlnet
        )
        self.pipeline.scheduler = UniPCMultistepScheduler.from_config(self.pipeline.scheduler.config)

        self.pipeline.to(device=device)
        if self.device.type == "cuda":
            self.pipeline.to(torch_dtype=torch.float16)

        self.is_initialized = True

    def __call__(self, image, negative_prompt="low quality, bad quality, deformed, low resolution", added_prompt=" , highest quality, highly realistic, very high resolution"):
        if not self.is_initialized:
            self.setup()

        initial_prompt = "super-hero character, best quality, extremely detailed"
        prompt = initial_prompt + prompt

        low_threshold = 100
        high_threshold = 200

        image = np.array(image)
        image = cv2.Canny(image, low_threshold, high_threshold)
        image = image[:, :, None]
        image = np.concatenate([image, image, image], axis=2)
        canny_image = Image.fromarray(image)

        return self.pipeline(
            prompt + added_prompt,
            canny_image,
            negative_prompt=negative_prompt,
            num_inference_steps=25,
        ).images[0]