modify model

ace_inference.py

@@ -1,356 +0,0 @@
-# -*- coding: utf-8 -*-
-# Copyright (c) Alibaba, Inc. and its affiliates.
-import copy
-import math
-import random
-
-import numpy as np
-import torch
-import torch.nn as nn
-import torch.nn.functional as F
-import torchvision.transforms.functional as TF
-from PIL import Image
-import torchvision.transforms as T
-from scepter.modules.model.registry import DIFFUSIONS
-from scepter.modules.model.utils.basic_utils import check_list_of_list
-from scepter.modules.model.utils.basic_utils import \
-    pack_imagelist_into_tensor_v2 as pack_imagelist_into_tensor
-from scepter.modules.model.utils.basic_utils import (
-    to_device, unpack_tensor_into_imagelist)
-from scepter.modules.utils.distribute import we
-from scepter.modules.utils.logger import get_logger
-
-from scepter.modules.inference.diffusion_inference import DiffusionInference, get_model
-
-
-def process_edit_image(images,
-                       masks,
-                       tasks,
-                       max_seq_len=1024,
-                       max_aspect_ratio=4,
-                       d=16,
-                       **kwargs):
-
-    if not isinstance(images, list):
-        images = [images]
-    if not isinstance(masks, list):
-        masks = [masks]
-    if not isinstance(tasks, list):
-        tasks = [tasks]
-
-    img_tensors = []
-    mask_tensors = []
-    for img, mask, task in zip(images, masks, tasks):
-        if mask is None or mask == '':
-            mask = Image.new('L', img.size, 0)
-        W, H = img.size
-        if H / W > max_aspect_ratio:
-            img = TF.center_crop(img, [int(max_aspect_ratio * W), W])
-            mask = TF.center_crop(mask, [int(max_aspect_ratio * W), W])
-        elif W / H > max_aspect_ratio:
-            img = TF.center_crop(img, [H, int(max_aspect_ratio * H)])
-            mask = TF.center_crop(mask, [H, int(max_aspect_ratio * H)])
-
-        H, W = img.height, img.width
-        scale = min(1.0, math.sqrt(max_seq_len / ((H / d) * (W / d))))
-        rH = int(H * scale) // d * d  # ensure divisible by self.d
-        rW = int(W * scale) // d * d
-
-        img = TF.resize(img, (rH, rW),
-                        interpolation=TF.InterpolationMode.BICUBIC)
-        mask = TF.resize(mask, (rH, rW),
-                         interpolation=TF.InterpolationMode.NEAREST_EXACT)
-
-        mask = np.asarray(mask)
-        mask = np.where(mask > 128, 1, 0)
-        mask = mask.astype(
-            np.float32) if np.any(mask) else np.ones_like(mask).astype(
-                np.float32)
-
-        img_tensor = TF.to_tensor(img).to(we.device_id)
-        img_tensor = TF.normalize(img_tensor,
-                                  mean=[0.5, 0.5, 0.5],
-                                  std=[0.5, 0.5, 0.5])
-        mask_tensor = TF.to_tensor(mask).to(we.device_id)
-        if task in ['inpainting', 'Try On', 'Inpainting']:
-            mask_indicator = mask_tensor.repeat(3, 1, 1)
-            img_tensor[mask_indicator == 1] = -1.0
-        img_tensors.append(img_tensor)
-        mask_tensors.append(mask_tensor)
-    return img_tensors, mask_tensors
-
-class TextEmbedding(nn.Module):
-    def __init__(self, embedding_shape):
-        super().__init__()
-        self.pos = nn.Parameter(data=torch.zeros(embedding_shape))
-
-class ACEFluxLCInference(DiffusionInference):
-    def __init__(self, logger=None):
-        if logger is None:
-            logger = get_logger(name='scepter')
-        self.logger = logger
-        self.loaded_model = {}
-        self.loaded_model_name = [
-            'diffusion_model', 'first_stage_model', 'cond_stage_model', 'ref_cond_stage_model'
-        ]
-
-    def init_from_cfg(self, cfg):
-        self.name = cfg.NAME
-        self.is_default = cfg.get('IS_DEFAULT', False)
-        self.use_dynamic_model = cfg.get('USE_DYNAMIC_MODEL', True)
-        module_paras = self.load_default(cfg.get('DEFAULT_PARAS', None))
-        assert cfg.have('MODEL')
-        self.size_factor = cfg.get('SIZE_FACTOR', 8)
-        self.diffusion_model = self.infer_model(
-            cfg.MODEL.DIFFUSION_MODEL, module_paras.get(
-                'DIFFUSION_MODEL',
-                None)) if cfg.MODEL.have('DIFFUSION_MODEL') else None
-        self.first_stage_model = self.infer_model(
-            cfg.MODEL.FIRST_STAGE_MODEL,
-            module_paras.get(
-                'FIRST_STAGE_MODEL',
-                None)) if cfg.MODEL.have('FIRST_STAGE_MODEL') else None
-        self.cond_stage_model = self.infer_model(
-            cfg.MODEL.COND_STAGE_MODEL,
-            module_paras.get(
-                'COND_STAGE_MODEL',
-                None)) if cfg.MODEL.have('COND_STAGE_MODEL') else None
-
-        self.ref_cond_stage_model = self.infer_model(
-            cfg.MODEL.REF_COND_STAGE_MODEL,
-            module_paras.get(
-                'REF_COND_STAGE_MODEL',
-                None)) if cfg.MODEL.have('REF_COND_STAGE_MODEL') else None
-
-        self.diffusion = DIFFUSIONS.build(cfg.MODEL.DIFFUSION,
-                                          logger=self.logger)
-        self.interpolate_func = lambda x: (F.interpolate(
-            x.unsqueeze(0),
-            scale_factor=1 / self.size_factor,
-            mode='nearest-exact') if x is not None else None)
-
-        self.max_seq_length = cfg.get("MAX_SEQ_LENGTH", 4096)
-        self.src_max_seq_length = cfg.get("SRC_MAX_SEQ_LENGTH", 1024)
-        self.image_token = cfg.MODEL.get("IMAGE_TOKEN", "<img>")
-
-        self.text_indentifers = cfg.MODEL.get('TEXT_IDENTIFIER', [])
-        self.use_text_pos_embeddings = cfg.MODEL.get('USE_TEXT_POS_EMBEDDINGS',
-                                                     False)
-        if self.use_text_pos_embeddings:
-            self.text_position_embeddings = TextEmbedding(
-                (10, 4096)).eval().requires_grad_(False).to(we.device_id)
-        else:
-            self.text_position_embeddings = None
-
-        if not self.use_dynamic_model:
-            self.dynamic_load(self.first_stage_model, 'first_stage_model')
-            self.dynamic_load(self.cond_stage_model, 'cond_stage_model')
-            if self.ref_cond_stage_model is not None: self.dynamic_load(self.ref_cond_stage_model, 'ref_cond_stage_model')
-            self.dynamic_load(self.diffusion_model, 'diffusion_model')
-
-    def upscale_resize(self, image, interpolation=T.InterpolationMode.BILINEAR):
-        c, H, W = image.shape
-        scale = max(1.0, math.sqrt(self.max_seq_length / ((H / 16) * (W / 16))))
-        rH = int(H * scale) // 16 * 16  # ensure divisible by self.d
-        rW = int(W * scale) // 16 * 16
-        image = T.Resize((rH, rW), interpolation=interpolation, antialias=True)(image)
-        return image
-
-
-    @torch.no_grad()
-    def encode_first_stage(self, x, **kwargs):
-        _, dtype = self.get_function_info(self.first_stage_model, 'encode')
-        with torch.autocast('cuda',
-                            enabled=dtype in ('float16', 'bfloat16'),
-                            dtype=getattr(torch, dtype)):
-            def run_one_image(u):
-                zu = get_model(self.first_stage_model).encode(u)
-                if isinstance(zu, (tuple, list)):
-                    zu = zu[0]
-                return zu
-
-            z = [run_one_image(u.unsqueeze(0) if u.dim() == 3 else u) for u in x]
-            return z
-
-
-    @torch.no_grad()
-    def decode_first_stage(self, z):
-        _, dtype = self.get_function_info(self.first_stage_model, 'decode')
-        with torch.autocast('cuda',
-                            enabled=dtype in ('float16', 'bfloat16'),
-                            dtype=getattr(torch, dtype)):
-            return [get_model(self.first_stage_model).decode(zu) for zu in z]
-
-    def noise_sample(self, num_samples, h, w, seed, device = None, dtype = torch.bfloat16):
-        noise = torch.randn(
-            num_samples,
-            16,
-            # allow for packing
-            2 * math.ceil(h / 16),
-            2 * math.ceil(w / 16),
-            device=device,
-            dtype=dtype,
-            generator=torch.Generator(device=device).manual_seed(seed),
-        )
-        return noise
-
-    # def preprocess_prompt(self, prompt):
-    #     prompt_ = [[pp] if isinstance(pp, str) else pp for pp in prompt]
-    #     for pp_id, pp in enumerate(prompt_):
-    #         prompt_[pp_id] = [""] + pp
-    #         for p_id, p in enumerate(prompt_[pp_id]):
-    #             prompt_[pp_id][p_id] = self.image_token + self.text_indentifers[p_id] + " " + p
-    #         prompt_[pp_id] = [f";".join(prompt_[pp_id])]
-    #     return prompt_
-
-    @torch.no_grad()
-    def __call__(self,
-                 image=None,
-                 mask=None,
-                 prompt='',
-                 task=None,
-                 negative_prompt='',
-                 output_height=1024,
-                 output_width=1024,
-                 sampler='flow_euler',
-                 sample_steps=20,
-                 guide_scale=3.5,
-                 seed=-1,
-                 history_io=None,
-                 tar_index=0,
-                 align=0,
-                 **kwargs):
-        input_image, input_mask = image, mask
-        seed = seed if seed >= 0 else random.randint(0, 2**32 - 1)
-        if input_image is not None:
-            # assert isinstance(input_image, list) and isinstance(input_mask, list)
-            if task is None:
-                task = [''] * len(input_image)
-            if not isinstance(prompt, list):
-                prompt = [prompt] * len(input_image)
-            prompt = [
-                pp.replace('{image}', f'{{image{i}}}') if i > 0 else pp
-                for i, pp in enumerate(prompt)
-            ]
-            edit_image, edit_image_mask = process_edit_image(
-                input_image, input_mask, task, max_seq_len=self.src_max_seq_length)
-            image, image_mask = self.upscale_resize(edit_image[tar_index]), self.upscale_resize(edit_image_mask[
-               tar_index])
-            # edit_image, edit_image_mask = [[self.upscale_resize(i) for i in edit_image]], [[self.upscale_resize(i) for i in edit_image_mask]]
-            # image, image_mask = edit_image[tar_index], edit_image_mask[tar_index]
-            edit_image, edit_image_mask = [edit_image], [edit_image_mask]
-        else:
-            edit_image = edit_image_mask = [[]]
-            image = torch.zeros(
-                size=[3, int(output_height),
-                      int(output_width)])
-            image_mask = torch.ones(
-                size=[1, int(output_height),
-                      int(output_width)])
-            if not isinstance(prompt, list):
-                prompt = [prompt]
-
-        image, image_mask, prompt = [image], [image_mask], [prompt],
-        align = [align for p in prompt] if isinstance(align, int) else align
-
-        assert check_list_of_list(prompt) and check_list_of_list(
-            edit_image) and check_list_of_list(edit_image_mask)
-        # negative prompt is not used
-        image = to_device(image)
-        ctx = {}
-        # Get Noise Shape
-        self.dynamic_load(self.first_stage_model, 'first_stage_model')
-        x = self.encode_first_stage(image)
-        self.dynamic_unload(self.first_stage_model,
-                            'first_stage_model',
-                            skip_loaded=not self.use_dynamic_model)
-
-        g = torch.Generator(device=we.device_id).manual_seed(seed)
-
-        noise = [
-            torch.randn((1, 16, i.shape[2], i.shape[3]), device=we.device_id, dtype=torch.bfloat16).normal_(generator=g)
-            for i in x
-        ]
-        noise, x_shapes = pack_imagelist_into_tensor(noise)
-        ctx['x_shapes'] = x_shapes
-        ctx['align'] = align
-
-        image_mask = to_device(image_mask, strict=False)
-        cond_mask = [self.interpolate_func(i) for i in image_mask
-                     ] if image_mask is not None else [None] * len(image)
-        ctx['x_mask'] = cond_mask
-        # Encode Prompt
-        instruction_prompt = [[pp[-1]] if "{image}" in pp[-1] else ["{image} " + pp[-1]] for pp in prompt]
-        self.dynamic_load(self.cond_stage_model, 'cond_stage_model')
-        function_name, dtype = self.get_function_info(self.cond_stage_model)
-        cont = getattr(get_model(self.cond_stage_model), function_name)(instruction_prompt)
-        cont["context"] = [ct[-1] for ct in cont["context"]]
-        cont["y"] = [ct[-1] for ct in cont["y"]]
-        self.dynamic_unload(self.cond_stage_model,
-                            'cond_stage_model',
-                            skip_loaded=not self.use_dynamic_model)
-        ctx.update(cont)
-
-        # Encode Edit Images
-        self.dynamic_load(self.first_stage_model, 'first_stage_model')
-        edit_image = [to_device(i, strict=False) for i in edit_image]
-        edit_image_mask = [to_device(i, strict=False) for i in edit_image_mask]
-        e_img, e_mask = [], []
-        for u, m in zip(edit_image, edit_image_mask):
-            if u is None:
-                continue
-            if m is None:
-                m = [None] * len(u)
-            e_img.append(self.encode_first_stage(u, **kwargs))
-            e_mask.append([self.interpolate_func(i) for i in m])
-        self.dynamic_unload(self.first_stage_model,
-                            'first_stage_model',
-                            skip_loaded=not self.use_dynamic_model)
-        ctx['edit_x'] = e_img
-        ctx['edit_mask'] = e_mask
-        # Encode Ref Images
-        if guide_scale is not None:
-            guide_scale = torch.full((noise.shape[0],), guide_scale, device=noise.device, dtype=noise.dtype)
-        else:
-            guide_scale = None
-
-        # Diffusion Process
-        self.dynamic_load(self.diffusion_model, 'diffusion_model')
-        function_name, dtype = self.get_function_info(self.diffusion_model)
-        with torch.autocast('cuda',
-                            enabled=dtype in ('float16', 'bfloat16'),
-                            dtype=getattr(torch, dtype)):
-            latent = self.diffusion.sample(
-                noise=noise,
-                sampler=sampler,
-                model=get_model(self.diffusion_model),
-                model_kwargs={
-                    "cond": ctx, "guidance": guide_scale, "gc_seg": -1
-                },
-                steps=sample_steps,
-                show_progress=True,
-                guide_scale=guide_scale,
-                return_intermediate=None,
-                reverse_scale=-1,
-                **kwargs).float()
-        if self.use_dynamic_model: self.dynamic_unload(self.diffusion_model,
-                            'diffusion_model',
-                            skip_loaded=not self.use_dynamic_model)
-
-        # Decode to Pixel Space
-        self.dynamic_load(self.first_stage_model, 'first_stage_model')
-        samples = unpack_tensor_into_imagelist(latent, x_shapes)
-        x_samples = self.decode_first_stage(samples)
-        self.dynamic_unload(self.first_stage_model,
-                            'first_stage_model',
-                            skip_loaded=not self.use_dynamic_model)
-        x_samples = [x.squeeze(0) for x in x_samples]
-
-        imgs = [
-            torch.clamp((x_i.float() + 1.0) / 2.0,
-                        min=0.0,
-                        max=1.0).squeeze(0).permute(1, 2, 0).cpu().numpy()
-            for x_i in x_samples
-        ]
-        imgs = [Image.fromarray((img * 255).astype(np.uint8)) for img in imgs]
-        return imgs

example.py

@@ -1,370 +0,0 @@
-# -*- coding: utf-8 -*-
-# Copyright (c) Alibaba, Inc. and its affiliates.
-import os
-from PIL import Image
-from scepter.modules.utils.file_system import FS
-
-
-def download_image(image, local_path=None):
-    if not FS.exists(local_path):
-        local_path = FS.get_from(image, local_path=local_path)
-        if local_path.split(".")[-1] in ['jpg', 'jpeg']:
-            im = Image.open(local_path).convert("RGB")
-            im.save(local_path, format='JPEG')
-    return local_path
-
-
-def get_examples(cache_dir):
-    print('Downloading Examples ...')
-    examples = [
-        [
-            'Facial Editing',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/e33edc106953.png?raw=true',
-                os.path.join(cache_dir, 'examples/e33edc106953.jpg')), None,
-            None, '{image} let the man smile', 6666
-        ],
-        [
-            'Facial Editing',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/5d2bcc91a3e9.png?raw=true',
-                os.path.join(cache_dir, 'examples/5d2bcc91a3e9.jpg')), None,
-            None, 'let the man in {image} wear sunglasses', 9999
-        ],
-        [
-            'Facial Editing',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/3a52eac708bd.png?raw=true',
-                os.path.join(cache_dir, 'examples/3a52eac708bd.jpg')), None,
-            None, '{image} red hair', 9999
-        ],
-        [
-            'Facial Editing',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/3f4dc464a0ea.png?raw=true',
-                os.path.join(cache_dir, 'examples/3f4dc464a0ea.jpg')), None,
-            None, '{image} let the man serious', 99999
-        ],
-        [
-            'Controllable Generation',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/131ca90fd2a9.png?raw=true',
-                os.path.join(cache_dir,
-                             'examples/131ca90fd2a9.jpg')), None, None,
-            '"A person sits contemplatively on the ground, surrounded by falling autumn leaves. Dressed in a green sweater and dark blue pants, they rest their chin on their hand, exuding a relaxed demeanor. Their stylish checkered slip-on shoes add a touch of flair, while a black purse lies in their lap. The backdrop of muted brown enhances the warm, cozy atmosphere of the scene." , generate the image that corresponds to the given scribble {image}.',
-            613725
-        ],
-        [
-            'Render Text',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/33e9f27c2c48.png?raw=true',
-                os.path.join(cache_dir, 'examples/33e9f27c2c48.jpg')),
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/33e9f27c2c48_mask.png?raw=true',
-                os.path.join(cache_dir,
-                             'examples/33e9f27c2c48_mask.jpg')), None,
-            'Put the text "C A T" at the position marked by mask in the {image}',
-            6666
-        ],
-        [
-            'Style Transfer',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/9e73e7eeef55.png?raw=true',
-                os.path.join(cache_dir, 'examples/9e73e7eeef55.jpg')), None,
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/2e02975293d6.png?raw=true',
-                os.path.join(cache_dir, 'examples/2e02975293d6.jpg')),
-            'edit {image} based on the style of {image1} ', 99999
-        ],
-        [
-            'Outpainting',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/f2b22c08be3f.png?raw=true',
-                os.path.join(cache_dir, 'examples/f2b22c08be3f.jpg')),
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/f2b22c08be3f_mask.png?raw=true',
-                os.path.join(cache_dir,
-                             'examples/f2b22c08be3f_mask.jpg')), None,
-            'Could the {image} be widened within the space designated by mask, while retaining the original?',
-            6666
-        ],
-        [
-            'Image Segmentation',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/db3ebaa81899.png?raw=true',
-                os.path.join(cache_dir, 'examples/db3ebaa81899.jpg')), None,
-            None, '{image} Segmentation', 6666
-        ],
-        [
-            'Depth Estimation',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/f1927c4692ba.png?raw=true',
-                os.path.join(cache_dir, 'examples/f1927c4692ba.jpg')), None,
-            None, '{image} Depth Estimation', 6666
-        ],
-        [
-            'Pose Estimation',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/014e5bf3b4d1.png?raw=true',
-                os.path.join(cache_dir, 'examples/014e5bf3b4d1.jpg')), None,
-            None, '{image} distinguish the poses of the figures', 999999
-        ],
-        [
-            'Scribble Extraction',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/5f59a202f8ac.png?raw=true',
-                os.path.join(cache_dir, 'examples/5f59a202f8ac.jpg')), None,
-            None, 'Generate a scribble of {image}, please.', 6666
-        ],
-        [
-            'Mosaic',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/3a2f52361eea.png?raw=true',
-                os.path.join(cache_dir, 'examples/3a2f52361eea.jpg')), None,
-            None, 'Adapt {image} into a mosaic representation.', 6666
-        ],
-        [
-            'Edge map Extraction',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/b9d1e519d6e5.png?raw=true',
-                os.path.join(cache_dir, 'examples/b9d1e519d6e5.jpg')), None,
-            None, 'Get the edge-enhanced result for {image}.', 6666
-        ],
-        [
-            'Grayscale',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/c4ebbe2ba29b.png?raw=true',
-                os.path.join(cache_dir, 'examples/c4ebbe2ba29b.jpg')), None,
-            None, 'transform {image} into a black and white one', 6666
-        ],
-        [
-            'Contour Extraction',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/19652d0f6c4b.png?raw=true',
-                os.path.join(cache_dir,
-                             'examples/19652d0f6c4b.jpg')), None, None,
-            'Would you be able to make a contour picture from {image} for me?',
-            6666
-        ],
-        [
-            'Controllable Generation',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/249cda2844b7.png?raw=true',
-                os.path.join(cache_dir,
-                             'examples/249cda2844b7.jpg')), None, None,
-            'Following the segmentation outcome in mask of {image}, develop a real-life image using the explanatory note in "a mighty cat lying on the bed”.',
-            6666
-        ],
-        [
-            'Controllable Generation',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/411f6c4b8e6c.png?raw=true',
-                os.path.join(cache_dir,
-                             'examples/411f6c4b8e6c.jpg')), None, None,
-            'use the depth map {image} and the text caption "a cut white cat" to create a corresponding graphic image',
-            999999
-        ],
-        [
-            'Controllable Generation',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/a35c96ed137a.png?raw=true',
-                os.path.join(cache_dir,
-                             'examples/a35c96ed137a.jpg')), None, None,
-            'help translate this posture schema {image} into a colored image based on the context I provided "A beautiful woman Climbing the climbing wall, wearing a harness and climbing gear, skillfully maneuvering up the wall with her back to the camera, with a safety rope."',
-            3599999
-        ],
-        [
-            'Controllable Generation',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/dcb2fc86f1ce.png?raw=true',
-                os.path.join(cache_dir,
-                             'examples/dcb2fc86f1ce.jpg')), None, None,
-            'Transform and generate an image using mosaic {image} and "Monarch butterflies gracefully perch on vibrant purple flowers, showcasing their striking orange and black wings in a lush garden setting." description',
-            6666
-        ],
-        [
-            'Controllable Generation',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/4cd4ee494962.png?raw=true',
-                os.path.join(cache_dir,
-                             'examples/4cd4ee494962.jpg')), None, None,
-            'make this {image} colorful as per the "beautiful sunflowers"',
-            6666
-        ],
-        [
-            'Controllable Generation',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/a47e3a9cd166.png?raw=true',
-                os.path.join(cache_dir,
-                             'examples/a47e3a9cd166.jpg')), None, None,
-            'Take the edge conscious {image} and the written guideline "A whimsical animated character is depicted holding a delectable cake adorned with blue and white frosting and a drizzle of chocolate. The character wears a yellow headband with a bow, matching a cozy yellow sweater. Her dark hair is styled in a braid, tied with a yellow ribbon. With a golden fork in hand, she stands ready to enjoy a slice, exuding an air of joyful anticipation. The scene is creatively rendered with a charming and playful aesthetic." and produce a realistic image.',
-            613725
-        ],
-        [
-            'Controllable Generation',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/d890ed8a3ac2.png?raw=true',
-                os.path.join(cache_dir,
-                             'examples/d890ed8a3ac2.jpg')), None, None,
-            'creating a vivid image based on {image} and description "This image features a delicious rectangular tart with a flaky, golden-brown crust. The tart is topped with evenly sliced tomatoes, layered over a creamy cheese filling. Aromatic herbs are sprinkled on top, adding a touch of green and enhancing the visual appeal. The background includes a soft, textured fabric and scattered white flowers, creating an elegant and inviting presentation. Bright red tomatoes in the upper right corner hint at the fresh ingredients used in the dish."',
-            6666
-        ],
-        [
-            'Image Denoising',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/0844a686a179.png?raw=true',
-                os.path.join(cache_dir,
-                             'examples/0844a686a179.jpg')), None, None,
-            'Eliminate noise interference in {image} and maximize the crispness to obtain superior high-definition quality',
-            6666
-        ],
-        [
-            'Inpainting',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/fa91b6b7e59b.png?raw=true',
-                os.path.join(cache_dir, 'examples/fa91b6b7e59b.jpg')),
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/fa91b6b7e59b_mask.png?raw=true',
-                os.path.join(cache_dir,
-                             'examples/fa91b6b7e59b_mask.jpg')), None,
-            'Ensure to overhaul the parts of the {image} indicated by the mask.',
-            6666
-        ],
-        [
-            'Inpainting',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/632899695b26.png?raw=true',
-                os.path.join(cache_dir, 'examples/632899695b26.jpg')),
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/632899695b26_mask.png?raw=true',
-                os.path.join(cache_dir,
-                             'examples/632899695b26_mask.jpg')), None,
-            'Refashion the mask portion of {image} in accordance with "A yellow egg with a smiling face painted on it"',
-            6666
-        ],
-        [
-            'General Editing',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/354d17594afe.png?raw=true',
-                os.path.join(cache_dir,
-                             'examples/354d17594afe.jpg')), None, None,
-            '{image} change the dog\'s posture to walking in the water, and change the background to green plants and a pond.',
-            6666
-        ],
-        [
-            'General Editing',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/38946455752b.png?raw=true',
-                os.path.join(cache_dir,
-                             'examples/38946455752b.jpg')), None, None,
-            '{image} change the color of the dress from white to red and the model\'s hair color red brown to blonde.Other parts remain unchanged',
-            6669
-        ],
-        [
-            'Facial Editing',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/3ba5202f0cd8.png?raw=true',
-                os.path.join(cache_dir,
-                             'examples/3ba5202f0cd8.jpg')), None, None,
-            'Keep the same facial feature in @3ba5202f0cd8, change the woman\'s clothing from a Blue denim jacket to a white turtleneck sweater and adjust her posture so that she is supporting her chin with both hands. Other aspects, such as background, hairstyle, facial expression, etc, remain unchanged.',
-            99999
-        ],
-        [
-            'Facial Editing',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/369365b94725.png?raw=true',
-                os.path.join(cache_dir, 'examples/369365b94725.jpg')), None,
-            None, '{image} Make her looking at the camera', 6666
-        ],
-        [
-            'Facial Editing',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/92751f2e4a0e.png?raw=true',
-                os.path.join(cache_dir, 'examples/92751f2e4a0e.jpg')), None,
-            None, '{image} Remove the smile from his face', 9899999
-        ],
-        [
-            'Remove Text',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/8530a6711b2e.png?raw=true',
-                os.path.join(cache_dir, 'examples/8530a6711b2e.jpg')), None,
-            None, 'Aim to remove any textual element in {image}', 6666
-        ],
-        [
-            'Remove Text',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/c4d7fb28f8f6.png?raw=true',
-                os.path.join(cache_dir, 'examples/c4d7fb28f8f6.jpg')),
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/c4d7fb28f8f6_mask.png?raw=true',
-                os.path.join(cache_dir,
-                             'examples/c4d7fb28f8f6_mask.jpg')), None,
-            'Rub out any text found in the mask sector of the {image}.', 6666
-        ],
-        [
-            'Remove Object',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/e2f318fa5e5b.png?raw=true',
-                os.path.join(cache_dir,
-                             'examples/e2f318fa5e5b.jpg')), None, None,
-            'Remove the unicorn in this {image}, ensuring a smooth edit.',
-            99999
-        ],
-        [
-            'Remove Object',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/1ae96d8aca00.png?raw=true',
-                os.path.join(cache_dir, 'examples/1ae96d8aca00.jpg')),
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/1ae96d8aca00_mask.png?raw=true',
-                os.path.join(cache_dir, 'examples/1ae96d8aca00_mask.jpg')),
-            None, 'Discard the contents of the mask area from {image}.', 99999
-        ],
-        [
-            'Add Object',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/80289f48e511.png?raw=true',
-                os.path.join(cache_dir, 'examples/80289f48e511.jpg')),
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/80289f48e511_mask.png?raw=true',
-                os.path.join(cache_dir,
-                             'examples/80289f48e511_mask.jpg')), None,
-            'add a Hot Air Balloon into the {image}, per the mask', 613725
-        ],
-        [
-            'Style Transfer',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/d725cb2009e8.png?raw=true',
-                os.path.join(cache_dir, 'examples/d725cb2009e8.jpg')), None,
-            None, 'Change the style of {image} to colored pencil style', 99999
-        ],
-        [
-            'Style Transfer',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/e0f48b3fd010.png?raw=true',
-                os.path.join(cache_dir, 'examples/e0f48b3fd010.jpg')), None,
-            None, 'make {image} to Walt Disney Animation style', 99999
-        ],
-        [
-            'Try On',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/ee4ca60b8c96.png?raw=true',
-                os.path.join(cache_dir, 'examples/ee4ca60b8c96.jpg')),
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/ee4ca60b8c96_mask.png?raw=true',
-                os.path.join(cache_dir, 'examples/ee4ca60b8c96_mask.jpg')),
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/ebe825bbfe3c.png?raw=true',
-                os.path.join(cache_dir, 'examples/ebe825bbfe3c.jpg')),
-            'Change the cloth in {image} to the one in {image1}', 99999
-        ],
-        [
-            'Workflow',
-            download_image(
-                'https://github.com/ali-vilab/ace-page/blob/main/assets/examples/cb85353c004b.png?raw=true',
-                os.path.join(cache_dir, 'examples/cb85353c004b.jpg')), None,
-            None, '<workflow> ice cream {image}', 99999
-        ],
-    ]
-    print('Finish. Start building UI ...')
-    return examples

utils.py

@@ -1,95 +0,0 @@
-#copyright (c) Alibaba, Inc. and its affiliates.
-import torch
-import torchvision.transforms as T
-from PIL import Image
-from torchvision.transforms.functional import InterpolationMode
-
-IMAGENET_MEAN = (0.485, 0.456, 0.406)
-IMAGENET_STD = (0.229, 0.224, 0.225)
-
-
-def build_transform(input_size):
-    MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
-    transform = T.Compose([
-        T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
-        T.Resize((input_size, input_size),
-                 interpolation=InterpolationMode.BICUBIC),
-        T.ToTensor(),
-        T.Normalize(mean=MEAN, std=STD)
-    ])
-    return transform
-
-
-def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height,
-                              image_size):
-    best_ratio_diff = float('inf')
-    best_ratio = (1, 1)
-    area = width * height
-    for ratio in target_ratios:
-        target_aspect_ratio = ratio[0] / ratio[1]
-        ratio_diff = abs(aspect_ratio - target_aspect_ratio)
-        if ratio_diff < best_ratio_diff:
-            best_ratio_diff = ratio_diff
-            best_ratio = ratio
-        elif ratio_diff == best_ratio_diff:
-            if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
-                best_ratio = ratio
-    return best_ratio
-
-
-def dynamic_preprocess(image,
-                       min_num=1,
-                       max_num=12,
-                       image_size=448,
-                       use_thumbnail=False):
-    orig_width, orig_height = image.size
-    aspect_ratio = orig_width / orig_height
-
-    # calculate the existing image aspect ratio
-    target_ratios = set((i, j) for n in range(min_num, max_num + 1)
-                        for i in range(1, n + 1) for j in range(1, n + 1)
-                        if i * j <= max_num and i * j >= min_num)
-    target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
-
-    # find the closest aspect ratio to the target
-    target_aspect_ratio = find_closest_aspect_ratio(aspect_ratio,
-                                                    target_ratios, orig_width,
-                                                    orig_height, image_size)
-
-    # calculate the target width and height
-    target_width = image_size * target_aspect_ratio[0]
-    target_height = image_size * target_aspect_ratio[1]
-    blocks = target_aspect_ratio[0] * target_aspect_ratio[1]
-
-    # resize the image
-    resized_img = image.resize((target_width, target_height))
-    processed_images = []
-    for i in range(blocks):
-        box = ((i % (target_width // image_size)) * image_size,
-               (i // (target_width // image_size)) * image_size,
-               ((i % (target_width // image_size)) + 1) * image_size,
-               ((i // (target_width // image_size)) + 1) * image_size)
-        # split the image
-        split_img = resized_img.crop(box)
-        processed_images.append(split_img)
-    assert len(processed_images) == blocks
-    if use_thumbnail and len(processed_images) != 1:
-        thumbnail_img = image.resize((image_size, image_size))
-        processed_images.append(thumbnail_img)
-    return processed_images
-
-
-def load_image(image_file, input_size=448, max_num=12):
-    if isinstance(image_file, str):
-        image = Image.open(image_file).convert('RGB')
-    else:
-        image = image_file
-    transform = build_transform(input_size=input_size)
-    images = dynamic_preprocess(image,
-                                image_size=input_size,
-                                use_thumbnail=True,
-                                max_num=max_num)
-    pixel_values = [transform(image) for image in images]
-    pixel_values = torch.stack(pixel_values)
-    return pixel_values
-