inference.py

from __future__ import annotations

import gc
import json
import pathlib
import sys

import gradio as gr
import PIL.Image
import torch
from diffusers import StableDiffusionInpaintPipeline


class InferencePipeline:
    def __init__(self):
        self.pipe = None
        self.device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

    def clear(self) -> None:
        del self.pipe
        self.pipe = None
        torch.cuda.empty_cache()
        gc.collect()

    def load_pipe(self, realfill_model: str) -> None:
        pipe = StableDiffusionInpaintPipeline.from_pretrained(
            realfill_model, torch_dtype=torch.float16
        ).to(self.device)
        pipe = pipe.to(self.device)
        self.pipe = pipe

    def run(
        self,
        realfill_model: str,
        target_image: PIL.Image,
        target_mask: PIL.Image,
        seed: int,
        n_steps: int,
        guidance_scale: float,
    ) -> PIL.Image.Image:
        if not torch.cuda.is_available():
            raise gr.Error("CUDA is not available.")

        self.load_pipe(realfill_model)

        image = PIL.Image.open(target_image)
        mask_image = PIL.Image.open(target_mask)

        generator = torch.Generator(device=self.device).manual_seed(seed)
        out = self.pipe(
            "a photo of sks",
            image=image, 
            mask_image=mask_image, 
            num_inference_steps=n_steps,
            guidance_scale=guidance_scale,
            generator=generator,
        ).images[0]  # type: ignore

        erode_kernel = PIL.ImageFilter.MaxFilter(3)
        mask_image = mask_image.filter(erode_kernel)

        result = PIL.Image.composite(result, out, mask_image)
        return result