handler.py

import base64
from io import BytesIO

import torch

from utils import pipes, tiling, wallpaper, images

default_desired_output_width = 1024
default_text2img_inference_steps = 50
default_text2img_guidance_scale = 7.14
default_img2img_inference_steps = 50
default_img2img_strength = 0.6


class EndpointHandler:
    def __init__(self, path=""):
        self.sdxl_pipe = pipes.create_stable_diffusion_xl_pipeline("cuda:0", enable_tiling=False)
        self.sdxl_img2img_pipe = pipes.create_stable_diffusion_xl_img2img_pipe("cuda:1")

    def __call__(self, data):
        self.sdxl_pipe.to("cuda:0")
        self.sdxl_img2img_pipe.to("cuda:1")

        prompt = data.pop("inputs", data)
        desired_output_width = int(data.get("desired_output_width", default_desired_output_width))
        text2img_inference_steps = int(data.get("text2img_inference_steps", default_text2img_inference_steps))
        img2img_inference_steps = int(data.get("img2img_inference_steps", default_img2img_inference_steps))
        text2img_guidance_scale = float(data.get("text2img_guidance_scale", default_text2img_guidance_scale))
        img2img_strength = float(data.get("img2img_strength", default_img2img_strength))

        print(
            f"Prompt: {prompt}, Desired Output Width: {desired_output_width}, Text2Img Inference Steps: {text2img_inference_steps}, Img2Img Inference Steps: {img2img_inference_steps}, Text2Img Guidance Scale: {text2img_guidance_scale}, Img2Img Strength: {img2img_strength}")

        original_image_size = tiling.compute_input_tile_width_for_desired_output(desired_output_width)

        original_image = \
            self.sdxl_pipe(prompt=prompt, num_inference_steps=text2img_inference_steps, width=original_image_size,
                           height=original_image_size, guidance_scale=text2img_guidance_scale).images[0]
        tiled_canvas = wallpaper.prepare_for_diamond_drop(original_image)
        inner_rotated_tile = images.extract_larger_inner_rotated_tile_from_image(tiled_canvas, original_image_size,
                                                                                 original_image_size)

        half_drop = wallpaper.convert_tile_to_half_drop(inner_rotated_tile)

        half_drop_buffer = BytesIO()
        half_drop.save(half_drop_buffer, format="PNG")
        half_drop_string = base64.b64encode(half_drop_buffer.getvalue())

        torch.cuda.empty_cache()

        return {"image": half_drop_string.decode()}