Merge branch 'dev' into dev_device_support

gen_img_diffusers.py

@@ -101,11 +101,14 @@ import tools.original_control_net as original_control_net
 from tools.original_control_net import ControlNetInfo
 from library.original_unet import UNet2DConditionModel, InferUNet2DConditionModel
 from library.original_unet import FlashAttentionFunction
+from library.utils import GradualLatent, EulerAncestralDiscreteSchedulerGL
 
 from XTI_hijack import unet_forward_XTI, downblock_forward_XTI, upblock_forward_XTI
-from library.utils import setup_logging
+from library.utils import setup_logging, add_logging_arguments
+
 setup_logging()
 import logging
+
 logger = logging.getLogger(__name__)
 
 # scheduler:
@@ -452,6 +455,8 @@ class PipelineLike:
         self.control_nets: List[ControlNetInfo] = []
         self.control_net_enabled = True  # control_netsが空ならTrueでもFalseでもControlNetは動作しない
 
+        self.gradual_latent: GradualLatent = None
+
     # Textual Inversion
     def add_token_replacement(self, target_token_id, rep_token_ids):
         self.token_replacements[target_token_id] = rep_token_ids
@@ -482,6 +487,14 @@ class PipelineLike:
     def set_control_nets(self, ctrl_nets):
         self.control_nets = ctrl_nets
 
+    def set_gradual_latent(self, gradual_latent):
+        if gradual_latent is None:
+            print("gradual_latent is disabled")
+            self.gradual_latent = None
+        else:
+            print(f"gradual_latent is enabled: {gradual_latent}")
+            self.gradual_latent = gradual_latent  # (ds_ratio, start_timesteps, every_n_steps, ratio_step)
+
     # region xformersとか使う部分：独自に書き換えるので関係なし
 
     def enable_xformers_memory_efficient_attention(self):
@@ -955,7 +968,49 @@ class PipelineLike:
         else:
             text_emb_last = text_embeddings
 
+        enable_gradual_latent = False
+        if self.gradual_latent:
+            if not hasattr(self.scheduler, "set_gradual_latent_params"):
+                print("gradual_latent is not supported for this scheduler. Ignoring.")
+                print(self.scheduler.__class__.__name__)
+            else:
+                enable_gradual_latent = True
+                step_elapsed = 1000
+                current_ratio = self.gradual_latent.ratio
+
+                # first, we downscale the latents to the specified ratio / 最初に指定された比率にlatentsをダウンスケールする
+                height, width = latents.shape[-2:]
+                org_dtype = latents.dtype
+                if org_dtype == torch.bfloat16:
+                    latents = latents.float()
+                latents = torch.nn.functional.interpolate(
+                    latents, scale_factor=current_ratio, mode="bicubic", align_corners=False
+                ).to(org_dtype)
+
+                # apply unsharp mask / アンシャープマスクを適用する
+                if self.gradual_latent.gaussian_blur_ksize:
+                    latents = self.gradual_latent.apply_unshark_mask(latents)
+
         for i, t in enumerate(tqdm(timesteps)):
+            resized_size = None
+            if enable_gradual_latent:
+                # gradually upscale the latents / latentsを徐々にアップスケールする
+                if (
+                    t < self.gradual_latent.start_timesteps
+                    and current_ratio < 1.0
+                    and step_elapsed >= self.gradual_latent.every_n_steps
+                ):
+                    current_ratio = min(current_ratio + self.gradual_latent.ratio_step, 1.0)
+                    # make divisible by 8 because size of latents must be divisible at bottom of UNet
+                    h = int(height * current_ratio) // 8 * 8
+                    w = int(width * current_ratio) // 8 * 8
+                    resized_size = (h, w)
+                    self.scheduler.set_gradual_latent_params(resized_size, self.gradual_latent)
+                    step_elapsed = 0
+                else:
+                    self.scheduler.set_gradual_latent_params(None, None)
+                step_elapsed += 1
+
             # expand the latents if we are doing classifier free guidance
             latent_model_input = latents.repeat((num_latent_input, 1, 1, 1))
             latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
@@ -1536,7 +1591,9 @@ class PipelineLike:
         image_embeddings = self.vgg16_feat_model(image)["feat"]
 
         # バッチサイズが複数だと正しく動くかわからない
-        loss = ((image_embeddings - guide_embeddings) ** 2).mean() * guidance_scale  # MSE style transferでコンテンツの損失はMSEなので
+        loss = (
+            (image_embeddings - guide_embeddings) ** 2
+        ).mean() * guidance_scale  # MSE style transferでコンテンツの損失はMSEなので
 
         grads = -torch.autograd.grad(loss, latents)[0]
         if isinstance(self.scheduler, LMSDiscreteScheduler):
@@ -2127,6 +2184,7 @@ class BatchDataBase(NamedTuple):
     mask_image: Any
     clip_prompt: str
     guide_image: Any
+    raw_prompt: str
 
 
 class BatchDataExt(NamedTuple):
@@ -2246,7 +2304,7 @@ def main(args):
         scheduler_cls = EulerDiscreteScheduler
         scheduler_module = diffusers.schedulers.scheduling_euler_discrete
     elif args.sampler == "euler_a" or args.sampler == "k_euler_a":
-        scheduler_cls = EulerAncestralDiscreteScheduler
+        scheduler_cls = EulerAncestralDiscreteSchedulerGL
         scheduler_module = diffusers.schedulers.scheduling_euler_ancestral_discrete
     elif args.sampler == "dpmsolver" or args.sampler == "dpmsolver++":
         scheduler_cls = DPMSolverMultistepScheduler
@@ -2509,6 +2567,29 @@ def main(args):
     if args.ds_depth_1 is not None:
         unet.set_deep_shrink(args.ds_depth_1, args.ds_timesteps_1, args.ds_depth_2, args.ds_timesteps_2, args.ds_ratio)
 
+    # Gradual Latent
+    if args.gradual_latent_timesteps is not None:
+        if args.gradual_latent_unsharp_params:
+            us_params = args.gradual_latent_unsharp_params.split(",")
+            us_ksize, us_sigma, us_strength = [float(v) for v in us_params[:3]]
+            us_target_x = True if len(us_params) <= 3 else bool(int(us_params[3]))
+            us_ksize = int(us_ksize)
+        else:
+            us_ksize, us_sigma, us_strength, us_target_x = None, None, None, None
+
+        gradual_latent = GradualLatent(
+            args.gradual_latent_ratio,
+            args.gradual_latent_timesteps,
+            args.gradual_latent_every_n_steps,
+            args.gradual_latent_ratio_step,
+            args.gradual_latent_s_noise,
+            us_ksize,
+            us_sigma,
+            us_strength,
+            us_target_x,
+        )
+        pipe.set_gradual_latent(gradual_latent)
+
     # Extended Textual Inversion および Textual Inversionを処理する
     if args.XTI_embeddings:
         diffusers.models.UNet2DConditionModel.forward = unet_forward_XTI
@@ -2530,7 +2611,9 @@ def main(args):
             embeds = next(iter(data.values()))
 
             if type(embeds) != torch.Tensor:
-                raise ValueError(f"weight file does not contains Tensor / 重みファイルのデータがTensorではありません: {embeds_file}")
+                raise ValueError(
+                    f"weight file does not contains Tensor / 重みファイルのデータがTensorではありません: {embeds_file}"
+                )
 
             num_vectors_per_token = embeds.size()[0]
             token_string = os.path.splitext(os.path.basename(embeds_file))[0]
@@ -2630,7 +2713,7 @@ def main(args):
         logger.info(f"reading prompts from {args.from_file}")
         with open(args.from_file, "r", encoding="utf-8") as f:
             prompt_list = f.read().splitlines()
-            prompt_list = [d for d in prompt_list if len(d.strip()) > 0]
+            prompt_list = [d for d in prompt_list if len(d.strip()) > 0 and d[0] != "#"]
     elif args.prompt is not None:
         prompt_list = [args.prompt]
     else:
@@ -2760,7 +2843,9 @@ def main(args):
 
         logger.info(f"loaded {len(guide_images)} guide images for guidance")
         if len(guide_images) == 0:
-            logger.info(f"No guide image, use previous generated image. / ガイド画像がありません。直前に生成した画像を使います: {args.image_path}")
+            logger.info(
+                f"No guide image, use previous generated image. / ガイド画像がありません。直前に生成した画像を使います: {args.image_path}"
+            )
             guide_images = None
     else:
         guide_images = None
@@ -2874,13 +2959,14 @@ def main(args):
             # このバッチの情報を取り出す
             (
                 return_latents,
-                (step_first, _, _, _, init_image, mask_image, _, guide_image),
+                (step_first, _, _, _, init_image, mask_image, _, guide_image, _),
                 (width, height, steps, scale, negative_scale, strength, network_muls, num_sub_prompts),
             ) = batch[0]
             noise_shape = (LATENT_CHANNELS, height // DOWNSAMPLING_FACTOR, width // DOWNSAMPLING_FACTOR)
 
             prompts = []
             negative_prompts = []
+            raw_prompts = []
             start_code = torch.zeros((batch_size, *noise_shape), device=device, dtype=dtype)
             noises = [
                 torch.zeros((batch_size, *noise_shape), device=device, dtype=dtype)
@@ -2911,11 +2997,16 @@ def main(args):
             all_images_are_same = True
             all_masks_are_same = True
             all_guide_images_are_same = True
-            for i, (_, (_, prompt, negative_prompt, seed, init_image, mask_image, clip_prompt, guide_image), _) in enumerate(batch):
+            for i, (
+                _,
+                (_, prompt, negative_prompt, seed, init_image, mask_image, clip_prompt, guide_image, raw_prompt),
+                _,
+            ) in enumerate(batch):
                 prompts.append(prompt)
                 negative_prompts.append(negative_prompt)
                 seeds.append(seed)
                 clip_prompts.append(clip_prompt)
+                raw_prompts.append(raw_prompt)
 
                 if init_image is not None:
                     init_images.append(init_image)
@@ -3007,8 +3098,8 @@ def main(args):
             # save image
             highres_prefix = ("0" if highres_1st else "1") if highres_fix else ""
             ts_str = time.strftime("%Y%m%d%H%M%S", time.localtime())
-            for i, (image, prompt, negative_prompts, seed, clip_prompt) in enumerate(
-                zip(images, prompts, negative_prompts, seeds, clip_prompts)
+            for i, (image, prompt, negative_prompts, seed, clip_prompt, raw_prompt) in enumerate(
+                zip(images, prompts, negative_prompts, seeds, clip_prompts, raw_prompts)
             ):
                 if highres_fix:
                     seed -= 1  # record original seed
@@ -3024,6 +3115,8 @@ def main(args):
                     metadata.add_text("negative-scale", str(negative_scale))
                 if clip_prompt is not None:
                     metadata.add_text("clip-prompt", clip_prompt)
+                if raw_prompt is not None:
+                    metadata.add_text("raw-prompt", raw_prompt)
 
                 if args.use_original_file_name and init_images is not None:
                     if type(init_images) is list:
@@ -3046,7 +3139,9 @@ def main(args):
                         cv2.waitKey()
                         cv2.destroyAllWindows()
                 except ImportError:
-                    logger.info("opencv-python is not installed, cannot preview / opencv-pythonがインストールされていないためプレビューできません")
+                    logger.info(
+                        "opencv-python is not installed, cannot preview / opencv-pythonがインストールされていないためプレビューできません"
+                    )
 
             return images
 
@@ -3101,6 +3196,14 @@ def main(args):
                     ds_timesteps_2 = args.ds_timesteps_2
                     ds_ratio = args.ds_ratio
 
+                    # Gradual Latent
+                    gl_timesteps = None  # means no override
+                    gl_ratio = args.gradual_latent_ratio
+                    gl_every_n_steps = args.gradual_latent_every_n_steps
+                    gl_ratio_step = args.gradual_latent_ratio_step
+                    gl_s_noise = args.gradual_latent_s_noise
+                    gl_unsharp_params = args.gradual_latent_unsharp_params
+
                     prompt_args = raw_prompt.strip().split(" --")
                     prompt = prompt_args[0]
                     logger.info(f"prompt {prompt_index+1}/{len(prompt_list)}: {prompt}")
@@ -3203,10 +3306,52 @@ def main(args):
                             m = re.match(r"dsr ([\d\.]+)", parg, re.IGNORECASE)
                             if m:  # deep shrink ratio
                                 ds_ratio = float(m.group(1))
-                                 ds_depth_1 = ds_depth_1 if ds_depth_1 is not None else -1  # -1 means override
+                                ds_depth_1 = ds_depth_1 if ds_depth_1 is not None else -1  # -1 means override
                                 logger.info(f"deep shrink ratio: {ds_ratio}")
                                 continue
 
+                            # Gradual Latent
+                            m = re.match(r"glt ([\d\.]+)", parg, re.IGNORECASE)
+                            if m:  # gradual latent timesteps
+                                gl_timesteps = int(m.group(1))
+                                print(f"gradual latent timesteps: {gl_timesteps}")
+                                continue
+
+                            m = re.match(r"glr ([\d\.]+)", parg, re.IGNORECASE)
+                            if m:  # gradual latent ratio
+                                gl_ratio = float(m.group(1))
+                                gl_timesteps = gl_timesteps if gl_timesteps is not None else -1  # -1 means override
+                                print(f"gradual latent ratio: {ds_ratio}")
+                                continue
+
+                            m = re.match(r"gle ([\d\.]+)", parg, re.IGNORECASE)
+                            if m:  # gradual latent every n steps
+                                gl_every_n_steps = int(m.group(1))
+                                gl_timesteps = gl_timesteps if gl_timesteps is not None else -1  # -1 means override
+                                print(f"gradual latent every n steps: {gl_every_n_steps}")
+                                continue
+
+                            m = re.match(r"gls ([\d\.]+)", parg, re.IGNORECASE)
+                            if m:  # gradual latent ratio step
+                                gl_ratio_step = float(m.group(1))
+                                gl_timesteps = gl_timesteps if gl_timesteps is not None else -1  # -1 means override
+                                print(f"gradual latent ratio step: {gl_ratio_step}")
+                                continue
+
+                            m = re.match(r"glsn ([\d\.]+)", parg, re.IGNORECASE)
+                            if m:  # gradual latent s noise
+                                gl_s_noise = float(m.group(1))
+                                gl_timesteps = gl_timesteps if gl_timesteps is not None else -1  # -1 means override
+                                print(f"gradual latent s noise: {gl_s_noise}")
+                                continue
+
+                            m = re.match(r"glus ([\d\.\-,]+)", parg, re.IGNORECASE)
+                            if m:  # gradual latent unsharp params
+                                gl_unsharp_params = m.group(1)
+                                gl_timesteps = gl_timesteps if gl_timesteps is not None else -1  # -1 means override
+                                print(f"gradual latent unsharp params: {gl_unsharp_params}")
+                                continue
+
                         except ValueError as ex:
                             logger.info(f"Exception in parsing / 解析エラー: {parg}")
                             logger.info(ex)
@@ -3217,6 +3362,31 @@ def main(args):
                         ds_depth_1 = args.ds_depth_1 or 3
                     unet.set_deep_shrink(ds_depth_1, ds_timesteps_1, ds_depth_2, ds_timesteps_2, ds_ratio)
 
+                # override Gradual Latent
+                if gl_timesteps is not None:
+                    if gl_timesteps < 0:
+                        gl_timesteps = args.gradual_latent_timesteps or 650
+                    if gl_unsharp_params is not None:
+                        unsharp_params = gl_unsharp_params.split(",")
+                        us_ksize, us_sigma, us_strength = [float(v) for v in unsharp_params[:3]]
+                        print(unsharp_params)
+                        us_target_x = True if len(unsharp_params) < 4 else bool(int(unsharp_params[3]))
+                        us_ksize = int(us_ksize)
+                    else:
+                        us_ksize, us_sigma, us_strength, us_target_x = None, None, None, None
+                    gradual_latent = GradualLatent(
+                        gl_ratio,
+                        gl_timesteps,
+                        gl_every_n_steps,
+                        gl_ratio_step,
+                        gl_s_noise,
+                        us_ksize,
+                        us_sigma,
+                        us_strength,
+                        us_target_x,
+                    )
+                    pipe.set_gradual_latent(gradual_latent)
+
                 # prepare seed
                 if seeds is not None:  # given in prompt
                     # 数が足りないなら前のをそのまま使う
@@ -3284,7 +3454,9 @@ def main(args):
 
                 b1 = BatchData(
                     False,
-                    BatchDataBase(global_step, prompt, negative_prompt, seed, init_image, mask_image, clip_prompt, guide_image),
+                    BatchDataBase(
+                        global_step, prompt, negative_prompt, seed, init_image, mask_image, clip_prompt, guide_image, raw_prompt
+                    ),
                     BatchDataExt(
                         width,
                         height,
@@ -3319,16 +3491,25 @@ def main(args):
 def setup_parser() -> argparse.ArgumentParser:
     parser = argparse.ArgumentParser()
 
-    parser.add_argument("--v2", action="store_true", help="load Stable Diffusion v2.0 model / Stable Diffusion 2.0のモデルを読み込む")
+    add_logging_arguments(parser)
+
+    parser.add_argument(
+        "--v2", action="store_true", help="load Stable Diffusion v2.0 model / Stable Diffusion 2.0のモデルを読み込む"
+    )
     parser.add_argument(
         "--v_parameterization", action="store_true", help="enable v-parameterization training / v-parameterization学習を有効にする"
     )
     parser.add_argument("--prompt", type=str, default=None, help="prompt / プロンプト")
     parser.add_argument(
-        "--from_file", type=str, default=None, help="if specified, load prompts from this file / 指定時はプロンプトをファイルから読み込む"
+        "--from_file",
+        type=str,
+        default=None,
+        help="if specified, load prompts from this file / 指定時はプロンプトをファイルから読み込む",
     )
     parser.add_argument(
-        "--interactive", action="store_true", help="interactive mode (generates one image) / 対話モード（生成される画像は1枚になります）"
+        "--interactive",
+        action="store_true",
+        help="interactive mode (generates one image) / 対話モード（生成される画像は1枚になります）",
     )
     parser.add_argument(
         "--no_preview", action="store_true", help="do not show generated image in interactive mode / 対話モードで画像を表示しない"
@@ -3340,7 +3521,9 @@ def setup_parser() -> argparse.ArgumentParser:
     parser.add_argument("--strength", type=float, default=None, help="img2img strength / img2img時のstrength")
     parser.add_argument("--images_per_prompt", type=int, default=1, help="number of images per prompt / プロンプトあたりの出力枚数")
     parser.add_argument("--outdir", type=str, default="outputs", help="dir to write results to / 生成画像の出力先")
-    parser.add_argument("--sequential_file_name", action="store_true", help="sequential output file name / 生成画像のファイル名を連番にする")
+    parser.add_argument(
+        "--sequential_file_name", action="store_true", help="sequential output file name / 生成画像のファイル名を連番にする"
+    )
     parser.add_argument(
         "--use_original_file_name",
         action="store_true",
@@ -3394,9 +3577,14 @@ def setup_parser() -> argparse.ArgumentParser:
         default=7.5,
         help="unconditional guidance scale: eps = eps(x, empty) + scale * (eps(x, cond) - eps(x, empty)) / guidance scale",
     )
-    parser.add_argument("--ckpt", type=str, default=None, help="path to checkpoint of model / モデルのcheckpointファイルまたはディレクトリ")
     parser.add_argument(
-        "--vae", type=str, default=None, help="path to checkpoint of vae to replace / VAEを入れ替える場合、VAEのcheckpointファイルまたはディレクトリ"
+        "--ckpt", type=str, default=None, help="path to checkpoint of model / モデルのcheckpointファイルまたはディレクトリ"
+    )
+    parser.add_argument(
+        "--vae",
+        type=str,
+        default=None,
+        help="path to checkpoint of vae to replace / VAEを入れ替える場合、VAEのcheckpointファイルまたはディレクトリ",
     )
     parser.add_argument(
         "--tokenizer_cache_dir",
@@ -3432,25 +3620,46 @@ def setup_parser() -> argparse.ArgumentParser:
         help="use xformers by diffusers (Hypernetworks doesn't work) / Diffusersでxformersを使用する（Hypernetwork利用不可）",
     )
     parser.add_argument(
-        "--opt_channels_last", action="store_true", help="set channels last option to model / モデルにchannels lastを指定し最適化する"
+        "--opt_channels_last",
+        action="store_true",
+        help="set channels last option to model / モデルにchannels lastを指定し最適化する",
     )
     parser.add_argument(
-        "--network_module", type=str, default=None, nargs="*", help="additional network module to use / 追加ネットワークを使う時そのモジュール名"
+        "--network_module",
+        type=str,
+        default=None,
+        nargs="*",
+        help="additional network module to use / 追加ネットワークを使う時そのモジュール名",
     )
     parser.add_argument(
         "--network_weights", type=str, default=None, nargs="*", help="additional network weights to load / 追加ネットワークの重み"
     )
-    parser.add_argument("--network_mul", type=float, default=None, nargs="*", help="additional network multiplier / 追加ネットワークの効果の倍率")
     parser.add_argument(
-        "--network_args", type=str, default=None, nargs="*", help="additional arguments for network (key=value) / ネットワークへの追加の引数"
+        "--network_mul", type=float, default=None, nargs="*", help="additional network multiplier / 追加ネットワークの効果の倍率"
     )
-    parser.add_argument("--network_show_meta", action="store_true", help="show metadata of network model / ネットワークモデルのメタデータを表示する")
     parser.add_argument(
-        "--network_merge_n_models", type=int, default=None, help="merge this number of networks / この数だけネットワークをマージする"
+        "--network_args",
+        type=str,
+        default=None,
+        nargs="*",
+        help="additional arguments for network (key=value) / ネットワークへの追加の引数",
     )
-    parser.add_argument("--network_merge", action="store_true", help="merge network weights to original model / ネットワークの重みをマージする")
     parser.add_argument(
-        "--network_pre_calc", action="store_true", help="pre-calculate network for generation / ネットワークのあらかじめ計算して生成する"
+        "--network_show_meta", action="store_true", help="show metadata of network model / ネットワークモデルのメタデータを表示する"
+    )
+    parser.add_argument(
+        "--network_merge_n_models",
+        type=int,
+        default=None,
+        help="merge this number of networks / この数だけネットワークをマージする",
+    )
+    parser.add_argument(
+        "--network_merge", action="store_true", help="merge network weights to original model / ネットワークの重みをマージする"
+    )
+    parser.add_argument(
+        "--network_pre_calc",
+        action="store_true",
+        help="pre-calculate network for generation / ネットワークのあらかじめ計算して生成する",
     )
     parser.add_argument(
         "--network_regional_mask_max_color_codes",
@@ -3472,7 +3681,9 @@ def setup_parser() -> argparse.ArgumentParser:
         nargs="*",
         help="Embeddings files of Extended Textual Inversion / Extended Textual Inversionのembeddings",
     )
-    parser.add_argument("--clip_skip", type=int, default=None, help="layer number from bottom to use in CLIP / CLIPの後ろからn層目の出力を使う")
+    parser.add_argument(
+        "--clip_skip", type=int, default=None, help="layer number from bottom to use in CLIP / CLIPの後ろからn層目の出力を使う"
+    )
     parser.add_argument(
         "--max_embeddings_multiples",
         type=int,
@@ -3513,7 +3724,10 @@ def setup_parser() -> argparse.ArgumentParser:
         help="enable highres fix, reso scale for 1st stage / highres fixを有効にして最初の解像度をこのscaleにする",
     )
     parser.add_argument(
-        "--highres_fix_steps", type=int, default=28, help="1st stage steps for highres fix / highres fixの最初のステージのステップ数"
+        "--highres_fix_steps",
+        type=int,
+        default=28,
+        help="1st stage steps for highres fix / highres fixの最初のステージのステップ数",
     )
     parser.add_argument(
         "--highres_fix_strength",
@@ -3522,7 +3736,9 @@ def setup_parser() -> argparse.ArgumentParser:
         help="1st stage img2img strength for highres fix / highres fixの最初のステージのimg2img時のstrength、省略時はstrengthと同じ",
     )
     parser.add_argument(
-        "--highres_fix_save_1st", action="store_true", help="save 1st stage images for highres fix / highres fixの最初のステージの画像を保存する"
+        "--highres_fix_save_1st",
+        action="store_true",
+        help="save 1st stage images for highres fix / highres fixの最初のステージの画像を保存する",
     )
     parser.add_argument(
         "--highres_fix_latents_upscaling",
@@ -3530,7 +3746,10 @@ def setup_parser() -> argparse.ArgumentParser:
         help="use latents upscaling for highres fix / highres fixでlatentで拡大する",
     )
     parser.add_argument(
-        "--highres_fix_upscaler", type=str, default=None, help="upscaler module for highres fix / highres fixで使うupscalerのモジュール名"
+        "--highres_fix_upscaler",
+        type=str,
+        default=None,
+        help="upscaler module for highres fix / highres fixで使うupscalerのモジュール名",
     )
     parser.add_argument(
         "--highres_fix_upscaler_args",
@@ -3545,14 +3764,21 @@ def setup_parser() -> argparse.ArgumentParser:
     )
 
     parser.add_argument(
-        "--negative_scale", type=float, default=None, help="set another guidance scale for negative prompt / ネガティブプロンプトのscaleを指定する"
+        "--negative_scale",
+        type=float,
+        default=None,
+        help="set another guidance scale for negative prompt / ネガティブプロンプトのscaleを指定する",
     )
 
     parser.add_argument(
         "--control_net_models", type=str, default=None, nargs="*", help="ControlNet models to use / 使用するControlNetのモデル名"
     )
     parser.add_argument(
-        "--control_net_preps", type=str, default=None, nargs="*", help="ControlNet preprocess to use / 使用するControlNetのプリプロセス名"
+        "--control_net_preps",
+        type=str,
+        default=None,
+        nargs="*",
+        help="ControlNet preprocess to use / 使用するControlNetのプリプロセス名",
     )
     parser.add_argument("--control_net_weights", type=float, default=None, nargs="*", help="ControlNet weights / ControlNetの重み")
     parser.add_argument(
@@ -3590,6 +3816,45 @@ def setup_parser() -> argparse.ArgumentParser:
         "--ds_ratio", type=float, default=0.5, help="Deep Shrink ratio for downsampling / Deep Shrinkのdownsampling比率"
     )
 
+    # gradual latent
+    parser.add_argument(
+        "--gradual_latent_timesteps",
+        type=int,
+        default=None,
+        help="enable Gradual Latent hires fix and apply upscaling from this timesteps / Gradual Latent hires fixをこのtimestepsで有効にし、このtimestepsからアップスケーリングを適用する",
+    )
+    parser.add_argument(
+        "--gradual_latent_ratio",
+        type=float,
+        default=0.5,
+        help=" this size ratio, 0.5 means 1/2 / Gradual Latent hires fixをこのサイズ比率で有効にする、0.5は1/2を意味する",
+    )
+    parser.add_argument(
+        "--gradual_latent_ratio_step",
+        type=float,
+        default=0.125,
+        help="step to increase ratio for Gradual Latent / Gradual Latentのratioをどのくらいずつ上げるか",
+    )
+    parser.add_argument(
+        "--gradual_latent_every_n_steps",
+        type=int,
+        default=3,
+        help="steps to increase size of latents every this steps for Gradual Latent / Gradual Latentでlatentsのサイズをこのステップごとに上げる",
+    )
+    parser.add_argument(
+        "--gradual_latent_s_noise",
+        type=float,
+        default=1.0,
+        help="s_noise for Gradual Latent / Gradual Latentのs_noise",
+    )
+    parser.add_argument(
+        "--gradual_latent_unsharp_params",
+        type=str,
+        default=None,
+        help="unsharp mask parameters for Gradual Latent: ksize, sigma, strength, target-x (1 means True). `3,0.5,0.5,1` or `3,1.0,1.0,0` is recommended /"
+        + " Gradual Latentのunsharp maskのパラメータ: ksize, sigma, strength, target-x. `3,0.5,0.5,1` または `3,1.0,1.0,0` が推奨",
+    )
+
     return parser
 
 
@@ -3597,4 +3862,5 @@ if __name__ == "__main__":
     parser = setup_parser()
 
     args = parser.parse_args()
+    setup_logging(args, reset=True)
     main(args)