acmyu/KeyframesAI

1import glob
2import os
3import torch
4from torch import nn
5from PIL import Image
6import numpy as np
7from diffusers import UniPCMultistepScheduler
8from src.models.stage2_inpaint_unet_2d_condition import Stage2_InapintUNet2DConditionModel
9
10from torchvision import transforms
11from diffusers.models.controlnet import ControlNetConditioningEmbedding
12from transformers import CLIPImageProcessor
13from transformers import Dinov2Model
14from diffusers import AutoencoderKL, DDPMScheduler, UNet2DConditionModel,ControlNetModel,DDIMScheduler
15from src.pipelines.PCDMs_pipeline import PCDMsPipeline
16from single_extract_pose import inference_pose
17
18
19class ImageProjModel(torch.nn.Module):
20    """SD model with image prompt"""
21    def __init__(self, in_dim, hidden_dim, out_dim, dropout = 0.):
22        super().__init__()
23
24        self.net = nn.Sequential(
25            nn.Linear(in_dim, hidden_dim),
26            nn.GELU(),
27            nn.Dropout(dropout),
28            nn.LayerNorm(hidden_dim),
29            nn.Linear(hidden_dim, out_dim),
30            nn.Dropout(dropout)
31        )
32
33    def forward(self, x):  
34        return self.net(x)
35        
36class SDModel(torch.nn.Module):
37    """SD model with image prompt"""
38    def __init__(self, unet) -> None:
39        super().__init__()
40        self.image_proj_model = ImageProjModel(in_dim=1536, hidden_dim=768, out_dim=1024).to(device).to(dtype=torch.float16)
41        self.unet = unet
42        self.pose_proj = ControlNetConditioningEmbedding(
43            conditioning_embedding_channels=320,
44            block_out_channels=(16, 32, 96, 256),
45            conditioning_channels=3).to(device).to(dtype=torch.float16)
46
47
48def image_grid(imgs, rows, cols):
49    assert len(imgs) == rows * cols
50    w, h = imgs[0].size
51    print(w, h)
52    grid = Image.new("RGB", size=(cols * w, rows * h))
53    grid_w, grid_h = grid.size
54
55    for i, img in enumerate(imgs):
56        grid.paste(img, box=(i % cols * w, i // cols * h))
57    return grid
58
59def load_mydict(model_ckpt_path):
60    model_sd = torch.load(model_ckpt_path, map_location="cpu")["module"]
61
62    image_proj_model_dict = {}
63    pose_proj_dict = {}
64    unet_dict = {}
65    for k in model_sd.keys():
66        if k.startswith("pose_proj"):
67            pose_proj_dict[k.replace("pose_proj.", "")] = model_sd[k]
68
69        elif k.startswith("image_proj_model"):
70            image_proj_model_dict[k.replace("image_proj_model.", "")] = model_sd[k]
71
72
73        elif k.startswith("unet"):
74            unet_dict[k.replace("unet.", "")] = model_sd[k]
75        else:
76            print(k)
77    return image_proj_model_dict, pose_proj_dict, unet_dict
78
79
80        
81device = "cuda"
82pretrained_model_name_or_path ="stabilityai/stable-diffusion-2-1-base"
83image_encoder_path = "facebook/dinov2-giant"
84#model_ckpt_path = "./pcdms_ckpt.pt"   # ckpt path
85model_ckpt_path = 'fine_tuned_pcdms.pt'
86
87
88clip_image_processor = CLIPImageProcessor()
89img_transform = transforms.Compose([
90    transforms.ToTensor(),
91    transforms.Normalize([0.5], [0.5]),
92])
93
94generator = torch.Generator(device=device).manual_seed(42)
95unet = Stage2_InapintUNet2DConditionModel.from_pretrained(pretrained_model_name_or_path, torch_dtype=torch.float16,subfolder="unet",in_channels=9, low_cpu_mem_usage=False, ignore_mismatched_sizes=True).to(device)
96vae = AutoencoderKL.from_pretrained(pretrained_model_name_or_path,subfolder="vae").to(device, dtype=torch.float16)
97image_encoder = Dinov2Model.from_pretrained(image_encoder_path).to(device, dtype=torch.float16)
98noise_scheduler = DDIMScheduler(
99    num_train_timesteps=1000,
100    beta_start=0.00085,
101    beta_end=0.012,
102    beta_schedule="scaled_linear",
103    clip_sample=False,
104    set_alpha_to_one=False,
105    steps_offset=1,
106)
107
108image_proj_model = ImageProjModel(in_dim=1536, hidden_dim=768, out_dim=1024).to(device).to(dtype=torch.float16)
109pose_proj_model = ControlNetConditioningEmbedding(
110    conditioning_embedding_channels=320,
111    block_out_channels=(16, 32, 96, 256),
112    conditioning_channels=3).to(device).to(dtype=torch.float16)
113
114"""
115# load weight
116image_proj_model_dict, pose_proj_dict, unet_dict = load_mydict(model_ckpt_path)
117image_proj_model.load_state_dict(image_proj_model_dict)
118pose_proj_model.load_state_dict(pose_proj_dict)
119unet.load_state_dict(unet_dict)
120"""
121
122# load models
123model_sd = torch.load(model_ckpt_path, map_location="cpu")
124image_proj_model.load_state_dict(model_sd.image_proj_model.state_dict())
125pose_proj_model.load_state_dict(model_sd.pose_proj.state_dict())
126unet.load_state_dict(model_sd.unet.state_dict())
127
128pipe = PCDMsPipeline.from_pretrained(pretrained_model_name_or_path, unet=unet,  torch_dtype=torch.float16, scheduler=noise_scheduler,feature_extractor=None,safety_checker=None).to(device)
129
130print('====================== model load finish ===================')
131
132
133
134
135
136
137num_samples = 1
138image_size = (512, 512)
139s_img_path = 'imgs/sm.png' # input image 1
140target_pose_img = 'imgs/pose.png' # input image 2
141
142
143s_img = Image.open(s_img_path).convert("RGB").resize(image_size, Image.BICUBIC)
144black_image = Image.new("RGB", s_img.size, (0, 0, 0)).resize(image_size, Image.BICUBIC)
145
146s_img_t_mask = Image.new("RGB", (s_img.width * 2, s_img.height))
147s_img_t_mask.paste(s_img, (0, 0))
148s_img_t_mask.paste(black_image, (s_img.width, 0))
149
150s_pose = inference_pose(s_img_path, image_size=(image_size[1], image_size[0])).resize(image_size, Image.BICUBIC)
151print('source image width: {}, height: {}'.format(s_pose.width, s_pose.height))
152t_pose = Image.open(target_pose_img).convert("RGB").resize((image_size), Image.BICUBIC)
153
154st_pose = Image.new("RGB", (s_pose.width * 2, s_pose.height))
155st_pose.paste(s_pose, (0, 0))
156st_pose.paste(t_pose, (s_pose.width, 0))
157
158
159clip_s_img = clip_image_processor(images=s_img, return_tensors="pt").pixel_values
160vae_image = torch.unsqueeze(img_transform(s_img_t_mask), 0)
161cond_st_pose = torch.unsqueeze(img_transform(st_pose), 0)
162
163mask1 = torch.ones((1, 1, int(image_size[0] / 8), int(image_size[1] / 8))).to(device, dtype=torch.float16)
164mask0 = torch.zeros((1, 1, int(image_size[0] / 8), int(image_size[1] / 8))).to(device, dtype=torch.float16)
165mask = torch.cat([mask1, mask0], dim=3)
166
167
168with torch.inference_mode():
169    cond_pose = pose_proj_model(cond_st_pose.to(dtype=torch.float16, device=device))
170    simg_mask_latents = pipe.vae.encode(vae_image.to(device, dtype=torch.float16)).latent_dist.sample()
171    simg_mask_latents = simg_mask_latents * 0.18215
172
173    images_embeds = image_encoder(clip_s_img.to(device, dtype=torch.float16)).last_hidden_state
174    image_prompt_embeds = image_proj_model(images_embeds)
175    uncond_image_prompt_embeds = image_proj_model(torch.zeros_like(images_embeds))
176
177bs_embed, seq_len, _ = image_prompt_embeds.shape
178image_prompt_embeds = image_prompt_embeds.repeat(1, num_samples, 1)
179image_prompt_embeds = image_prompt_embeds.view(bs_embed * num_samples, seq_len, -1)
180uncond_image_prompt_embeds = uncond_image_prompt_embeds.repeat(1, num_samples, 1)
181uncond_image_prompt_embeds = uncond_image_prompt_embeds.view(bs_embed * num_samples, seq_len, -1)
182
183output, _ = pipe(
184    simg_mask_latents= simg_mask_latents,
185    mask = mask,
186    cond_pose = cond_pose,
187    prompt_embeds=image_prompt_embeds,
188    negative_prompt_embeds=uncond_image_prompt_embeds,
189    height=image_size[1],
190    width=image_size[0]*2,
191    num_images_per_prompt=num_samples,
192    guidance_scale=2.0,
193    generator=generator,
194    num_inference_steps=50,
195)
196
197output = output.images[-1]
198
199save_output = []
200result = output.crop((image_size[0], 0, image_size[0] * 2, image_size[1]))
201save_output.append(result.resize((352, 512), Image.BICUBIC))
202save_output.insert(0, t_pose.resize((352, 512), Image.BICUBIC))
203save_output.insert(0, s_img.resize((352, 512), Image.BICUBIC))
204grid = image_grid(save_output, 1, 3)
205grid.save("out.png")
206
207