File size: 5,465 Bytes
a9658c4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import re
import torch

from modules.processing import StableDiffusionProcessing, Processed

from scripts.animatediff_logger import logger_animatediff as logger
from scripts.animatediff_infotext import write_params_txt


class AnimateDiffPromptSchedule:

    def __init__(self):
        self.prompt_map = None
        self.original_prompt = None


    def save_infotext_img(self, p: StableDiffusionProcessing):
        if self.prompt_map is not None:
            p.prompts = [self.original_prompt for _ in range(p.batch_size)]


    def save_infotext_txt(self, res: Processed):
        if self.prompt_map is not None:
            parts = res.info.split('\nNegative prompt: ', 1)
            if len(parts) > 1:
                res.info = f"{self.original_prompt}\nNegative prompt: {parts[1]}"
                for i in range(len(res.infotexts)):
                    parts = res.infotexts[i].split('\nNegative prompt: ', 1)
                    if len(parts) > 1:
                        res.infotexts[i] = f"{self.original_prompt}\nNegative prompt: {parts[1]}"
                write_params_txt(res.info)


    def parse_prompt(self, p: StableDiffusionProcessing):
        if type(p.prompt) is not str:
            logger.warn("prompt is not str, cannot support prompt map")
            return

        lines = p.prompt.strip().split('\n')
        data = {
            'head_prompts': [],
            'mapp_prompts': {},
            'tail_prompts': []
        }

        mode = 'head'
        for line in lines:
            if mode == 'head':
                if re.match(r'^\d+:', line):
                    mode = 'mapp'
                else:
                    data['head_prompts'].append(line)
                    
            if mode == 'mapp':
                match = re.match(r'^(\d+): (.+)$', line)
                if match:
                    frame, prompt = match.groups()
                    data['mapp_prompts'][int(frame)] = prompt
                else:
                    mode = 'tail'
                    
            if mode == 'tail':
                data['tail_prompts'].append(line)
        
        if data['mapp_prompts']:
            logger.info("You are using prompt travel.")
            self.prompt_map = {}
            prompt_list = []
            last_frame = 0
            current_prompt = ''
            for frame, prompt in data['mapp_prompts'].items():
                prompt_list += [current_prompt for _ in range(last_frame, frame)]
                last_frame = frame
                current_prompt = f"{', '.join(data['head_prompts'])}, {prompt}, {', '.join(data['tail_prompts'])}"
                self.prompt_map[frame] = current_prompt
            prompt_list += [current_prompt for _ in range(last_frame, p.batch_size)]
            assert len(prompt_list) == p.batch_size, f"prompt_list length {len(prompt_list)} != batch_size {p.batch_size}"
            self.original_prompt = p.prompt
            p.prompt = prompt_list * p.n_iter


    def single_cond(self, center_frame, video_length: int, cond: torch.Tensor, closed_loop = False):
        if closed_loop:
            key_prev = list(self.prompt_map.keys())[-1]
            key_next = list(self.prompt_map.keys())[0]
        else:
            key_prev = list(self.prompt_map.keys())[0]
            key_next = list(self.prompt_map.keys())[-1]

        for p in self.prompt_map.keys():
            if p > center_frame:
                key_next = p
                break
            key_prev = p

        dist_prev = center_frame - key_prev
        if dist_prev < 0:
            dist_prev += video_length
        dist_next = key_next - center_frame
        if dist_next < 0:
            dist_next += video_length

        if key_prev == key_next or dist_prev + dist_next == 0:
            return cond[key_prev] if isinstance(cond, torch.Tensor) else {k: v[key_prev] for k, v in cond.items()}

        rate = dist_prev / (dist_prev + dist_next)
        if isinstance(cond, torch.Tensor):
            return AnimateDiffPromptSchedule.slerp(cond[key_prev], cond[key_next], rate)
        else: # isinstance(cond, dict)
            return {
                k: AnimateDiffPromptSchedule.slerp(v[key_prev], v[key_next], rate)
                for k, v in cond.items()
            }
    

    def multi_cond(self, cond: torch.Tensor, closed_loop = False):
        if self.prompt_map is None:
            return cond
        cond_list = [] if isinstance(cond, torch.Tensor) else {k: [] for k in cond.keys()}
        for i in range(cond.shape[0]):
            single_cond = self.single_cond(i, cond.shape[0], cond, closed_loop)
            if isinstance(cond, torch.Tensor):
                cond_list.append(single_cond)
            else:
                for k, v in single_cond.items():
                    cond_list[k].append(v)
        if isinstance(cond, torch.Tensor):
            return torch.stack(cond_list).to(cond.dtype).to(cond.device)
        else:
            return {k: torch.stack(v).to(cond[k].dtype).to(cond[k].device) for k, v in cond_list.items()}


    @staticmethod
    def slerp(
        v0: torch.Tensor, v1: torch.Tensor, t: float, DOT_THRESHOLD: float = 0.9995
    ) -> torch.Tensor:
        u0 = v0 / v0.norm()
        u1 = v1 / v1.norm()
        dot = (u0 * u1).sum()
        if dot.abs() > DOT_THRESHOLD:
            return (1.0 - t) * v0 + t * v1
        omega = dot.acos()
        return (((1.0 - t) * omega).sin() * v0 + (t * omega).sin() * v1) / omega.sin()