onnx-web/api/onnx_web/chain/upscale_outpaint.py

from logging import getLogger
from typing import Callable, List, Optional, Tuple

import numpy as np
import torch
from PIL import Image, ImageDraw, ImageOps

from ..diffusers.load import load_pipeline
from ..diffusers.utils import get_latents_from_seed, get_tile_latents, parse_prompt
from ..image import expand_image, mask_filter_none, noise_source_histogram
from ..output import save_image
from ..params import Border, ImageParams, Size, SizeChart, StageParams
from ..server import ServerContext
from ..utils import is_debug
from ..worker import ProgressCallback, WorkerContext
from .stage import BaseStage
from .tile import complete_tile, process_tile_grid, process_tile_order

logger = getLogger(__name__)


class UpscaleOutpaintStage(BaseStage):
    max_tile = SizeChart.unlimited

    def run(
        self,
        job: WorkerContext,
        server: ServerContext,
        stage: StageParams,
        params: ImageParams,
        sources: List[Image.Image],
        *,
        border: Border,
        stage_source: Optional[Image.Image] = None,
        stage_mask: Optional[Image.Image] = None,
        fill_color: str = "white",
        mask_filter: Callable = mask_filter_none,
        noise_source: Callable = noise_source_histogram,
        callback: Optional[ProgressCallback] = None,
        **kwargs,
    ) -> List[Image.Image]:
        _prompt_pairs, loras, inversions = parse_prompt(params)

        pipe_type = params.get_valid_pipeline("inpaint", params.pipeline)
        pipe = load_pipeline(
            server,
            params,
            pipe_type,
            job.get_device(),
            inversions=inversions,
            loras=loras,
        )

        outputs = []
        for source in sources:
            logger.info(
                "upscaling %s x %s image by expanding borders: %s",
                source.width,
                source.height,
                border,
            )

            margin_x = float(max(border.left, border.right))
            margin_y = float(max(border.top, border.bottom))
            overlap = min(margin_x / source.width, margin_y / source.height)

            if stage_mask is None:
                # if no mask was provided, keep the full source image
                stage_mask = Image.new("RGB", source.size, "black")

            # masks start as 512x512, resize to cover the source, then trim the extra
            mask_max = max(source.width, source.height)
            stage_mask = ImageOps.contain(stage_mask, (mask_max, mask_max))
            stage_mask = stage_mask.crop((0, 0, source.width, source.height))

            source, stage_mask, noise, full_size = expand_image(
                source,
                stage_mask,
                border,
                fill=fill_color,
                noise_source=noise_source,
                mask_filter=mask_filter,
            )

            full_latents = get_latents_from_seed(params.seed, Size(*full_size))

            draw_mask = ImageDraw.Draw(stage_mask)

            if is_debug():
                save_image(server, "last-source.png", source)
                save_image(server, "last-mask.png", stage_mask)
                save_image(server, "last-noise.png", noise)

            def outpaint(tile_source: Image.Image, dims: Tuple[int, int, int]):
                left, top, tile = dims
                size = Size(*tile_source.size)
                tile_mask = stage_mask.crop((left, top, left + tile, top + tile))
                tile_mask = complete_tile(tile_mask, tile)

                if is_debug():
                    save_image(server, "tile-source.png", tile_source)
                    save_image(server, "tile-mask.png", tile_mask)

                latents = get_tile_latents(full_latents, dims, size)
                if params.lpw():
                    logger.debug("using LPW pipeline for inpaint")
                    rng = torch.manual_seed(params.seed)
                    result = pipe.inpaint(
                        tile_source,
                        tile_mask,
                        params.prompt,
                        generator=rng,
                        guidance_scale=params.cfg,
                        height=size.height,
                        latents=latents,
                        negative_prompt=params.negative_prompt,
                        num_inference_steps=params.steps,
                        width=size.width,
                        callback=callback,
                    )
                else:
                    rng = np.random.RandomState(params.seed)
                    result = pipe(
                        params.prompt,
                        tile_source,
                        tile_mask,
                        height=size.height,
                        width=size.width,
                        num_inference_steps=params.steps,
                        guidance_scale=params.cfg,
                        negative_prompt=params.negative_prompt,
                        generator=rng,
                        latents=latents,
                        callback=callback,
                    )

                # once part of the image has been drawn, keep it
                draw_mask.rectangle((left, top, left + tile, top + tile), fill="black")
                return result.images[0]

            if params.pipeline == "panorama":
                logger.debug("outpainting with one shot panorama, no tiling")
                output = outpaint(source, (0, 0, max(source.width, source.height)))
            if overlap == 0:
                logger.debug("outpainting with 0 margin, using grid tiling")
                output = process_tile_grid(source, SizeChart.auto, 1, [outpaint])
            elif border.left == border.right and border.top == border.bottom:
                logger.debug(
                    "outpainting with an even border, using spiral tiling with %s overlap",
                    overlap,
                )
                output = process_tile_order(
                    stage.tile_order,
                    source,
                    SizeChart.auto,
                    1,
                    [outpaint],
                    overlap=overlap,
                )
            else:
                logger.debug("outpainting with an uneven border, using grid tiling")
                output = process_tile_grid(source, SizeChart.auto, 1, [outpaint])

            logger.info("final output image size: %sx%s", output.width, output.height)
            outputs.append(output)

        return outputs