GradualLatent highres fixのgen_img_diffusers.pyへの差分

patch_for_gen_img_diffusers.py

        # how much to increase the scale at each step: .125 seems to work well (because it's 1/8?)
        # 各ステップに拡大率をどのくらい増やすか：.125がよさそう（たぶん1/8なので）
        scale_step = 0.125

        # timesteps at which to start increasing the scale: model and prompt dependent
        # 拡大を開始するtimesteps:モデルとプロンプトによる
        start_timesteps = 800

        # how many steps to wait before increasing the scale again: smaller values lead to more artifacts, also depends on the total number of steps
        # 何ステップごとに拡大するか：総ステップ数にも関係する
        every_n_steps = 6

        # inp = input("scale step:")
        # try:
        #     scale_step = float(inp)
        # except:
        #     pass
        # inp = input("start timesteps:")
        # try:
        #     start_timesteps = int(inp)
        # except:
        #     pass
        # inp = input("every n steps:")
        # try:
        #     every_n_steps = int(inp)
        # except:
        #     pass

        # first, we downscale the latents to the half of the size
        # 最初に1/2に縮小する
        current_scale = 0.5
        height, width = latents.shape[-2:]
        latents = torch.nn.functional.interpolate(
            latents.float(), scale_factor=current_scale, mode="bicubic", align_corners=False
        ).to(latents.dtype)

        for i, t in enumerate(tqdm(timesteps)):
            # print(i, t, current_scale)
            if t < start_timesteps and current_scale < 1.0 and i % every_n_steps == 0:
                current_scale = min(current_scale + scale_step, 1.0)
                print(f"upscale at {i} step, scale={current_scale}")

                latents = torch.nn.functional.interpolate(
                    latents.float(),
                    size=(int(height * current_scale), int(width * current_scale)),
                    mode="bicubic",
                    align_corners=False,
                    # antialias=True,
                ).to(latents.dtype)
                steps_count = 0