File size: 7,335 Bytes
f961e67
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from functools import partial
from typing import Callable, List, Optional, Union, Tuple

import torch
from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer

from diffusers.models import AutoencoderKL, UNet2DConditionModel
# from diffusers import StableDiffusionPipeline
from diffusers.pipelines.stable_diffusion.safety_checker import \
    StableDiffusionSafetyChecker
from diffusers.schedulers import DDIMScheduler,PNDMScheduler, LMSDiscreteScheduler

from modified_stable_diffusion import ModifiedStableDiffusionPipeline
from torchvision.transforms import ToPILImage
import matplotlib.pyplot as plt



### credit to: https://github.com/cccntu/efficient-prompt-to-prompt
def backward_ddim(x_t, alpha_t, alpha_tm1, eps_xt):
    """ from noise to image"""
    return (
        alpha_tm1**0.5
        * (
            (alpha_t**-0.5 - alpha_tm1**-0.5) * x_t
            + ((1 / alpha_tm1 - 1) ** 0.5 - (1 / alpha_t - 1) ** 0.5) * eps_xt
        )
        + x_t
    )

def forward_ddim(x_t, alpha_t, alpha_tp1, eps_xt):
    """ from image to noise, it's the same as backward_ddim"""
    return backward_ddim(x_t, alpha_t, alpha_tp1, eps_xt)


class InversableStableDiffusionPipeline(ModifiedStableDiffusionPipeline):
    def __init__(self,
        vae,
        text_encoder,
        tokenizer,
        unet,
        scheduler,
        safety_checker,
        feature_extractor,
        requires_safety_checker: bool = False,
    ):

        super(InversableStableDiffusionPipeline, self).__init__(vae,
                text_encoder,
                tokenizer,
                unet,
                scheduler,
                safety_checker,
                feature_extractor,
                requires_safety_checker)

        self.forward_diffusion = partial(self.backward_diffusion, reverse_process=True)
        self.count = 0
    
    def get_random_latents(self, latents=None, height=512, width=512, generator=None):
        height = height or self.unet.config.sample_size * self.vae_scale_factor
        width = width or self.unet.config.sample_size * self.vae_scale_factor

        batch_size = 1
        device = self._execution_device

        num_channels_latents = self.unet.in_channels

        latents = self.prepare_latents(
            batch_size,
            num_channels_latents,
            height,
            width,
            self.text_encoder.dtype,
            device,
            generator,
            latents,
        )

        return latents

    @torch.inference_mode()
    def get_text_embedding(self, prompt):
        text_input_ids = self.tokenizer(
            prompt,
            padding="max_length",
            truncation=True,
            max_length=self.tokenizer.model_max_length,
            return_tensors="pt",
        ).input_ids
        text_embeddings = self.text_encoder(text_input_ids.to(self.device))[0]
        return text_embeddings
    
    @torch.inference_mode()
    def get_image_latents(self, image, sample=True, rng_generator=None):
        encoding_dist = self.vae.encode(image).latent_dist
        if sample:
            encoding = encoding_dist.sample(generator=rng_generator)
        else:
            encoding = encoding_dist.mode()
        latents = encoding * 0.18215
        return latents


    @torch.inference_mode()
    def backward_diffusion(
        self,
        use_old_emb_i=25,
        text_embeddings=None,
        old_text_embeddings=None,
        new_text_embeddings=None,
        latents: Optional[torch.FloatTensor] = None,
        num_inference_steps: int = 50,
        guidance_scale: float = 7.5,
        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
        callback_steps: Optional[int] = 1,
        reverse_process: True = False,
        **kwargs,
    ):
        """ Generate image from text prompt and latents
        """
        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
        # corresponds to doing no classifier free guidance.
        do_classifier_free_guidance = guidance_scale > 1.0
        # set timesteps
        self.scheduler.set_timesteps(num_inference_steps)
        # Some schedulers like PNDM have timesteps as arrays
        # It's more optimized to move all timesteps to correct device beforehand
        timesteps_tensor = self.scheduler.timesteps.to(self.device)
        # scale the initial noise by the standard deviation required by the scheduler
        latents = latents * self.scheduler.init_noise_sigma

        if old_text_embeddings is not None and new_text_embeddings is not None:
            prompt_to_prompt = True
        else:
            prompt_to_prompt = False


        for i, t in enumerate(self.progress_bar(timesteps_tensor if not reverse_process else reversed(timesteps_tensor))):
            if prompt_to_prompt:
                if i < use_old_emb_i:
                    text_embeddings = old_text_embeddings
                else:
                    text_embeddings = new_text_embeddings

            # expand the latents if we are doing classifier free guidance
            latent_model_input = (
                torch.cat([latents] * 2) if do_classifier_free_guidance else latents
            )
            latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)

            # predict the noise residual
            noise_pred = self.unet(
                latent_model_input, t, encoder_hidden_states=text_embeddings
            ).sample

            # perform guidance
            if do_classifier_free_guidance:
                noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
                noise_pred = noise_pred_uncond + guidance_scale * (
                    noise_pred_text - noise_pred_uncond
                )

            prev_timestep = (
                t
                - self.scheduler.config.num_train_timesteps
                // self.scheduler.num_inference_steps
            )
            # call the callback, if provided
            if callback is not None and i % callback_steps == 0:
                callback(i, t, latents)
            
            # ddim
            alpha_prod_t = self.scheduler.alphas_cumprod[t]
            alpha_prod_t_prev = (
                self.scheduler.alphas_cumprod[prev_timestep]
                if prev_timestep >= 0
                else self.scheduler.final_alpha_cumprod
            )
            if reverse_process:
                alpha_prod_t, alpha_prod_t_prev = alpha_prod_t_prev, alpha_prod_t
            latents = backward_ddim(
                x_t=latents,
                alpha_t=alpha_prod_t,
                alpha_tm1=alpha_prod_t_prev,
                eps_xt=noise_pred,
            )
        return latents

    
    @torch.inference_mode()
    def decode_image(self, latents: torch.FloatTensor, **kwargs):
        scaled_latents = 1 / 0.18215 * latents
        image = [
            self.vae.decode(scaled_latents[i : i + 1]).sample for i in range(len(latents))
        ]
        image = torch.cat(image, dim=0)
        return image

    @torch.inference_mode()
    def torch_to_numpy(self, image):
        image = (image / 2 + 0.5).clamp(0, 1)
        image = image.cpu().permute(0, 2, 3, 1).numpy()
        return image