File size: 10,130 Bytes
8ed2f16
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
import sys
import os
import json
import zipfile
import argparse

sys.path.append(os.getcwd())

from os.path import join as opj

import numpy as np
import torch
from PIL import Image, ImageFile
from torchvision.transforms import ToPILImage

ImageFile.LOAD_TRUNCATED_IMAGES = True

import transformers
from diffusers import (
    StableDiffusionControlNetImg2ImgPipeline,
    ControlNetModel,
    DPMSolverMultistepScheduler,
    AutoencoderKL,
)

from landmark_generation import generate_annotation
from natsort import ns, natsorted


class DomainImageGeneration:
    def __init__(
            self, src_path, args, model_name,
            device="cuda:0", clip_skip=2,
            use_anime_vae=False, save_path='',
            only_lora=False, model_base_path='./diffusion_model'
    ):
        """Initialize DomainImageGeneration class"""
        self.model_base_path = model_base_path
        self.device = device
        self.args = args
        self.src_path = src_path
        self.use_model = model_name
        self.only_lora = only_lora

        self.out_path_base = os.path.join(save_path, model_name)
        os.makedirs(self.out_path_base, exist_ok=True)

        self.diffusion_checkpoint_path = self._get_model_path(model_name)
        self.pipe = self._load_pipeline(model_name, use_anime_vae, clip_skip)
        print("All models loaded successfully")

    def _get_model_path(self, model_name):
        """Retrieve the model checkpoint path based on the model name"""
        base_path = self.model_base_path
        if model_name == "stable-diffusion-2-1-base":
            return os.path.join(base_path, "stable-diffusion-2-1-base")
        elif self.only_lora:
            return os.path.join(base_path, "stable-diffusion-v1-5")
        else:
            return os.path.join(base_path, model_name)

    def _load_controlnet(self, model_name):
        """Load the ControlNet model"""
        controlnet_path = os.path.join(self.model_base_path, 'ControlNetMediaPipeFace')
        if model_name == "stable-diffusion-2-1-base":
            controlnet_path += "old"

        return ControlNetModel.from_pretrained(
            controlnet_path, torch_dtype=torch.float16
        )

    def _load_pipeline(self, model_name, use_anime_vae, clip_skip):
        """Load the Stable Diffusion ControlNet Img2Img Pipeline"""
        controlnet = self._load_controlnet(model_name)

        if use_anime_vae:
            print("Using Anime VAE")
            anime_vae = AutoencoderKL.from_pretrained(
                "/nas8/liuhongyu/model/kl-f8-anime2", torch_dtype=torch.float16
            )
            pipeline = StableDiffusionControlNetImg2ImgPipeline.from_pretrained(
                self.diffusion_checkpoint_path, torch_dtype=torch.float16, safety_checker=None,
                vae=anime_vae, controlnet=controlnet
            ).to(self.device)

            self._load_lora(pipeline, "detail-tweaker-lora/add_detail.safetensors")

        elif model_name == "stable-diffusion-2-1-base":
            pipeline = StableDiffusionControlNetImg2ImgPipeline.from_pretrained(
                self.diffusion_checkpoint_path, torch_dtype=torch.float16,
                use_safetensors=True, controlnet=controlnet, variant="fp16"
            ).to(self.device)

        else:
            text_encoder = transformers.CLIPTextModel.from_pretrained(
                self.diffusion_checkpoint_path,
                subfolder="text_encoder",
                num_hidden_layers=12 - (clip_skip - 1),
                torch_dtype=torch.float16
            )
            pipeline = StableDiffusionControlNetImg2ImgPipeline.from_pretrained(
                self.diffusion_checkpoint_path, torch_dtype=torch.float16,
                use_safetensors=True, text_encoder=text_encoder,
                controlnet=controlnet, variant="fp16"
            ).to(self.device)

        self._apply_negative_embedding(pipeline, model_name)
        pipeline.scheduler = DPMSolverMultistepScheduler.from_config(
            pipeline.scheduler.config, use_karras_sigmas=True
        )

        print("Target diffusion model loaded")
        return pipeline

    def _load_lora(self, pipeline, lora_name):
        """Load LoRA weights into the model"""
        lora_path = f"/nas8/liuhongyu/model/{lora_name}"
        state_dict, network_alphas = pipeline.lora_state_dict(lora_path)
        pipeline.load_lora_into_unet(state_dict, network_alphas=network_alphas, unet=pipeline.unet)

    def _apply_negative_embedding(self, pipeline, model_name):
        """Apply negative embedding (textual inversion)"""
        if model_name not in ["stable-diffusion-xl-base-1.0", "stable-diffusion-2-1-base"]:
            if self.only_lora:
                self._load_lora(pipeline, model_name)
                pipeline.safety_checker = lambda images, clip_input: (images, None)
            else:
                pipeline.load_textual_inversion(
                    "/nas8/liuhongyu/lora_model",
                    weight_name="EasyNegativeV2.safetensors",
                    token="EasyNegative"
                )

    def image_generation(self, prompt,  strength=0.7,
                         guidance_scale=7.5, num_inference_steps=30):
        """Generate images using the diffusion model"""
        out_path = os.path.join(self.out_path_base, prompt.replace(" ", "_"))
        os.makedirs(out_path, exist_ok=True)
        src_img_list = natsorted(
            [f for f in os.listdir(self.src_path) if f.lower().endswith(('.jpg', '.jpeg', '.png', '.bmp', '.gif'))],
            alg=ns.PATH
        )

        all_gen_nums = 0  # Counter for generated images

        for img_name in src_img_list:
            src_img_pil = Image.open(os.path.join(self.src_path, img_name))
            control_image = generate_annotation(src_img_pil, max_faces=1)

            if control_image is not None:
                prompt_input = prompt

                # Apply different generation methods based on the model type
                if self.use_model in ['stable-diffusion-xl-base-1.0', 'stable-diffusion-2-1-base']:
                    trg_img_pil = self.pipe(
                        prompt=prompt_input,
                        image=src_img_pil,
                        strength=strength,
                        control_image=Image.fromarray(control_image),
                        guidance_scale=guidance_scale,
                        negative_prompt='worst quality, normal quality, low quality, low res, blurry',
                        num_inference_steps=num_inference_steps,
                        controlnet_conditioning_scale=1.5
                    )['images'][0]
                else:
                    trg_img_pil = self.pipe(
                        prompt=prompt_input,
                        image=src_img_pil,
                        control_image=Image.fromarray(control_image),
                        strength=strength,
                        guidance_scale=guidance_scale,
                        num_inference_steps=num_inference_steps,
                        controlnet_conditioning_scale=1.5,
                        negative_prompt='EasyNegative, worst quality, normal quality, low quality, low res, blurry'
                    )['images'][0]

                # Save the generated image if valid
                if np.array(trg_img_pil).max() > 0:
                    trg_img_pil.save(opj(out_path, img_name))
                    all_gen_nums += 1


def parse_args():
    """Parse command-line arguments."""
    parser = argparse.ArgumentParser(description="Domain Image Generation")

    parser.add_argument(
        "--src_img_path",
        type=str,
        default="demo_input",
        help="Path to the source image directory"
    )
    parser.add_argument("--strength", type=float, default=0.6, help="Strength of the sdeedit")
    parser.add_argument("--prompt", type=str, default=None, help="Text prompt for image generation")
    parser.add_argument("--guidance_scale", type=float, default=7.5, help="Guidance scale for Stable Diffusion")
    parser.add_argument("--sd_model_id", type=str, default="stable-diffusion-2-1-base", help="Stable Diffusion model ID")
    parser.add_argument("--num_inference_steps", type=int, default=30, help="Number of inference steps")
    parser.add_argument("--save_base", type=str, default="./output", help="Output directory for generated images")
    parser.add_argument("--device", type=str, default="cuda:0", help="Device to run inference on (e.g., 'cuda:0')")
    parser.add_argument("--seed", type=int, default=42, help="Random seed for reproducibility")
    parser.add_argument("--use_anime_vae", action="store_true", help="Enable Anime VAE for image generation")
    parser.add_argument("--model_base_path", type=str, default="./diffusion_model", help="Output directory for generated images")

    return parser.parse_args()


def set_random_seed(seed: int):
    """Set random seed for reproducibility."""
    torch.manual_seed(seed)
    np.random.seed(seed)
    if torch.cuda.is_available():
        torch.cuda.manual_seed_all(seed)


def main():
    """Main function to execute the image generation process."""
    args = parse_args()

    # Set random seed to ensure reproducibility
    set_random_seed(args.seed)

    # Check if the source image path exists
    if not os.path.exists(args.src_img_path):
        raise FileNotFoundError(f"❌ Source image path does not exist: {args.src_img_path}")

    # Ensure the output directory exists
    os.makedirs(args.save_base, exist_ok=True)

    # Initialize the DomainImageGeneration class and generate images
    data_generation = DomainImageGeneration(
        src_path=args.src_img_path,
        args=args,
        model_name=args.sd_model_id,
        save_path=args.save_base,
        device=args.device,
        use_anime_vae=args.use_anime_vae,
        model_base_path = args.model_base_path
    )

    # Start image generation
    data_generation.image_generation(prompt=args.prompt, strength=args.strength, guidance_scale=args.guidance_scale,num_inference_steps=args.num_inference_steps )


if __name__ == "__main__":
    main()