Spaces:
Paused
Paused
Upload folder using huggingface_hub
Browse files
app.py
ADDED
@@ -0,0 +1,136 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from __future__ import annotations
|
2 |
+
|
3 |
+
import gradio as gr
|
4 |
+
import spaces
|
5 |
+
from PIL import Image
|
6 |
+
import torch
|
7 |
+
|
8 |
+
from my_run import run as run_model
|
9 |
+
|
10 |
+
|
11 |
+
DESCRIPTION = '''# Turbo Edit
|
12 |
+
'''
|
13 |
+
|
14 |
+
@spaces.GPU
|
15 |
+
def main_pipeline(
|
16 |
+
input_image: str,
|
17 |
+
src_prompt: str,
|
18 |
+
tgt_prompt: str,
|
19 |
+
seed: int,
|
20 |
+
w1: float,
|
21 |
+
# w2: float,
|
22 |
+
):
|
23 |
+
|
24 |
+
w2 = 1.0
|
25 |
+
res_image = run_model(input_image, src_prompt, tgt_prompt, seed, w1, w2)
|
26 |
+
|
27 |
+
return res_image
|
28 |
+
|
29 |
+
|
30 |
+
with gr.Blocks(css='app/style.css') as demo:
|
31 |
+
gr.Markdown(DESCRIPTION)
|
32 |
+
|
33 |
+
gr.HTML(
|
34 |
+
'''<a href="https://huggingface.co/spaces/garibida/ReNoise-Inversion?duplicate=true">
|
35 |
+
<img src="https://bit.ly/3gLdBN6" alt="Duplicate Space"></a>Duplicate the Space to run privately without waiting in queue''')
|
36 |
+
|
37 |
+
with gr.Row():
|
38 |
+
with gr.Column():
|
39 |
+
input_image = gr.Image(
|
40 |
+
label="Input image",
|
41 |
+
type="filepath",
|
42 |
+
height=512,
|
43 |
+
width=512
|
44 |
+
)
|
45 |
+
src_prompt = gr.Text(
|
46 |
+
label='Source Prompt',
|
47 |
+
max_lines=1,
|
48 |
+
placeholder='Source Prompt',
|
49 |
+
)
|
50 |
+
tgt_prompt = gr.Text(
|
51 |
+
label='Target Prompt',
|
52 |
+
max_lines=1,
|
53 |
+
placeholder='Target Prompt',
|
54 |
+
)
|
55 |
+
with gr.Accordion("Advanced Options", open=False):
|
56 |
+
seed = gr.Slider(
|
57 |
+
label='seed',
|
58 |
+
minimum=0,
|
59 |
+
maximum=16*1024,
|
60 |
+
value=7865,
|
61 |
+
step=1
|
62 |
+
)
|
63 |
+
w1 = gr.Slider(
|
64 |
+
label='w',
|
65 |
+
minimum=1.0,
|
66 |
+
maximum=3.0,
|
67 |
+
value=1.5,
|
68 |
+
step=0.05
|
69 |
+
)
|
70 |
+
# w2 = gr.Slider(
|
71 |
+
# label='w2',
|
72 |
+
# minimum=1.0,
|
73 |
+
# maximum=3.0,
|
74 |
+
# value=1.0,
|
75 |
+
# step=0.05
|
76 |
+
# )
|
77 |
+
|
78 |
+
run_button = gr.Button('Edit')
|
79 |
+
with gr.Column():
|
80 |
+
# result = gr.Gallery(label='Result')
|
81 |
+
result = gr.Image(
|
82 |
+
label="Result",
|
83 |
+
type="pil",
|
84 |
+
height=512,
|
85 |
+
width=512
|
86 |
+
)
|
87 |
+
|
88 |
+
examples = [
|
89 |
+
[
|
90 |
+
"demo_im/WhatsApp Image 2024-05-17 at 17.32.53.jpeg", #input_image
|
91 |
+
"a painting of a white cat sleeping on a lotus flower", #src_prompt
|
92 |
+
"a painting of a white cat sleeping on a lotus flower", #tgt_prompt
|
93 |
+
4759, #seed
|
94 |
+
1.0, #w1
|
95 |
+
# 1.1, #w2
|
96 |
+
],
|
97 |
+
[
|
98 |
+
"demo_im/pexels-pixabay-458976.less.png", #input_image
|
99 |
+
"a squirrel standing in the grass", #src_prompt
|
100 |
+
"a squirrel standing in the grass", #tgt_prompt
|
101 |
+
6128, #seed
|
102 |
+
1.25, #w1
|
103 |
+
# 1.1, #w2
|
104 |
+
],
|
105 |
+
]
|
106 |
+
|
107 |
+
gr.Examples(examples=examples,
|
108 |
+
inputs=[
|
109 |
+
input_image,
|
110 |
+
src_prompt,
|
111 |
+
tgt_prompt,
|
112 |
+
seed,
|
113 |
+
w1,
|
114 |
+
# w2,
|
115 |
+
],
|
116 |
+
outputs=[
|
117 |
+
result
|
118 |
+
],
|
119 |
+
fn=main_pipeline,
|
120 |
+
cache_examples=True)
|
121 |
+
|
122 |
+
|
123 |
+
inputs = [
|
124 |
+
input_image,
|
125 |
+
src_prompt,
|
126 |
+
tgt_prompt,
|
127 |
+
seed,
|
128 |
+
w1,
|
129 |
+
# w2,
|
130 |
+
]
|
131 |
+
outputs = [
|
132 |
+
result
|
133 |
+
]
|
134 |
+
run_button.click(fn=main_pipeline, inputs=inputs, outputs=outputs)
|
135 |
+
|
136 |
+
demo.queue(max_size=50).launch(share=True, max_threads=100)
|
config.py
ADDED
@@ -0,0 +1,140 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from ml_collections import config_dict
|
2 |
+
import yaml
|
3 |
+
from diffusers.schedulers import (
|
4 |
+
DDIMScheduler,
|
5 |
+
EulerAncestralDiscreteScheduler,
|
6 |
+
EulerDiscreteScheduler,
|
7 |
+
DDPMScheduler,
|
8 |
+
)
|
9 |
+
from utils import (
|
10 |
+
deterministic_ddim_step,
|
11 |
+
deterministic_ddpm_step,
|
12 |
+
deterministic_euler_step,
|
13 |
+
deterministic_non_ancestral_euler_step,
|
14 |
+
)
|
15 |
+
|
16 |
+
BREAKDOWNS = ["x_t_c_hat", "x_t_hat_c", "no_breakdown", "x_t_hat_c_with_zeros"]
|
17 |
+
SCHEDULERS = ["ddpm", "ddim", "euler", "euler_non_ancestral"]
|
18 |
+
MODELS = [
|
19 |
+
"stabilityai/sdxl-turbo",
|
20 |
+
"stabilityai/stable-diffusion-xl-base-1.0",
|
21 |
+
"CompVis/stable-diffusion-v1-4",
|
22 |
+
]
|
23 |
+
|
24 |
+
def get_num_steps_actual(cfg):
|
25 |
+
return (
|
26 |
+
cfg.num_steps_inversion
|
27 |
+
- cfg.step_start
|
28 |
+
+ (1 if cfg.clean_step_timestep > 0 else 0)
|
29 |
+
if cfg.timesteps is None
|
30 |
+
else len(cfg.timesteps) + (1 if cfg.clean_step_timestep > 0 else 0)
|
31 |
+
)
|
32 |
+
|
33 |
+
|
34 |
+
def get_config(args):
|
35 |
+
if args.config_from_file and args.config_from_file != "":
|
36 |
+
with open(args.config_from_file, "r") as f:
|
37 |
+
cfg = config_dict.ConfigDict(yaml.safe_load(f))
|
38 |
+
|
39 |
+
num_steps_actual = get_num_steps_actual(cfg)
|
40 |
+
|
41 |
+
else:
|
42 |
+
cfg = config_dict.ConfigDict()
|
43 |
+
|
44 |
+
cfg.seed = 2
|
45 |
+
cfg.self_r = 0.5
|
46 |
+
cfg.cross_r = 0.9
|
47 |
+
cfg.eta = 1
|
48 |
+
cfg.scheduler_type = SCHEDULERS[0]
|
49 |
+
|
50 |
+
cfg.num_steps_inversion = 50 # timesteps: 999, 799, 599, 399, 199
|
51 |
+
cfg.step_start = 20
|
52 |
+
cfg.timesteps = None
|
53 |
+
cfg.noise_timesteps = None
|
54 |
+
num_steps_actual = get_num_steps_actual(cfg)
|
55 |
+
cfg.ws1 = [2] * num_steps_actual
|
56 |
+
cfg.ws2 = [1] * num_steps_actual
|
57 |
+
cfg.real_cfg_scale = 0
|
58 |
+
cfg.real_cfg_scale_save = 0
|
59 |
+
cfg.breakdown = BREAKDOWNS[1]
|
60 |
+
cfg.noise_shift_delta = 1
|
61 |
+
cfg.max_norm_zs = [-1] * (num_steps_actual - 1) + [15.5]
|
62 |
+
|
63 |
+
cfg.clean_step_timestep = 0
|
64 |
+
|
65 |
+
cfg.model = MODELS[1]
|
66 |
+
|
67 |
+
if cfg.scheduler_type == "ddim":
|
68 |
+
cfg.scheduler_class = DDIMScheduler
|
69 |
+
cfg.step_function = deterministic_ddim_step
|
70 |
+
elif cfg.scheduler_type == "ddpm":
|
71 |
+
cfg.scheduler_class = DDPMScheduler
|
72 |
+
cfg.step_function = deterministic_ddpm_step
|
73 |
+
elif cfg.scheduler_type == "euler":
|
74 |
+
cfg.scheduler_class = EulerAncestralDiscreteScheduler
|
75 |
+
cfg.step_function = deterministic_euler_step
|
76 |
+
elif cfg.scheduler_type == "euler_non_ancestral":
|
77 |
+
cfg.scheduler_class = EulerDiscreteScheduler
|
78 |
+
cfg.step_function = deterministic_non_ancestral_euler_step
|
79 |
+
else:
|
80 |
+
raise ValueError(f"Unknown scheduler type: {cfg.scheduler_type}")
|
81 |
+
|
82 |
+
with cfg.ignore_type():
|
83 |
+
if isinstance(cfg.max_norm_zs, (int, float)):
|
84 |
+
cfg.max_norm_zs = [cfg.max_norm_zs] * num_steps_actual
|
85 |
+
|
86 |
+
if isinstance(cfg.ws1, (int, float)):
|
87 |
+
cfg.ws1 = [cfg.ws1] * num_steps_actual
|
88 |
+
|
89 |
+
if isinstance(cfg.ws2, (int, float)):
|
90 |
+
cfg.ws2 = [cfg.ws2] * num_steps_actual
|
91 |
+
|
92 |
+
if not hasattr(cfg, "update_eta"):
|
93 |
+
cfg.update_eta = False
|
94 |
+
|
95 |
+
if not hasattr(cfg, "save_timesteps"):
|
96 |
+
cfg.save_timesteps = None
|
97 |
+
|
98 |
+
if not hasattr(cfg, "scheduler_timesteps"):
|
99 |
+
cfg.scheduler_timesteps = None
|
100 |
+
|
101 |
+
assert (
|
102 |
+
cfg.scheduler_type == "ddpm" or cfg.timesteps is None
|
103 |
+
), "timesteps must be None for ddim/euler"
|
104 |
+
|
105 |
+
assert (
|
106 |
+
len(cfg.max_norm_zs) == num_steps_actual
|
107 |
+
), f"len(cfg.max_norm_zs) ({len(cfg.max_norm_zs)}) != num_steps_actual ({num_steps_actual})"
|
108 |
+
|
109 |
+
assert (
|
110 |
+
len(cfg.ws1) == num_steps_actual
|
111 |
+
), f"len(cfg.ws1) ({len(cfg.ws1)}) != num_steps_actual ({num_steps_actual})"
|
112 |
+
|
113 |
+
assert (
|
114 |
+
len(cfg.ws2) == num_steps_actual
|
115 |
+
), f"len(cfg.ws2) ({len(cfg.ws2)}) != num_steps_actual ({num_steps_actual})"
|
116 |
+
|
117 |
+
assert cfg.noise_timesteps is None or len(cfg.noise_timesteps) == (
|
118 |
+
num_steps_actual - (1 if cfg.clean_step_timestep > 0 else 0)
|
119 |
+
), f"len(cfg.noise_timesteps) ({len(cfg.noise_timesteps)}) != num_steps_actual ({num_steps_actual})"
|
120 |
+
|
121 |
+
assert cfg.save_timesteps is None or len(cfg.save_timesteps) == (
|
122 |
+
num_steps_actual - (1 if cfg.clean_step_timestep > 0 else 0)
|
123 |
+
), f"len(cfg.save_timesteps) ({len(cfg.save_timesteps)}) != num_steps_actual ({num_steps_actual})"
|
124 |
+
|
125 |
+
return cfg
|
126 |
+
|
127 |
+
|
128 |
+
def get_config_name(config, args):
|
129 |
+
if args.folder_name is not None and args.folder_name != "":
|
130 |
+
return args.folder_name
|
131 |
+
timesteps_str = (
|
132 |
+
f"step_start {config.step_start}"
|
133 |
+
if config.timesteps is None
|
134 |
+
else f"timesteps {config.timesteps}"
|
135 |
+
)
|
136 |
+
return f"""\
|
137 |
+
ws1 {config.ws1[0]} ws2 {config.ws2[0]} real_cfg_scale {config.real_cfg_scale} {timesteps_str} \
|
138 |
+
real_cfg_scale_save {config.real_cfg_scale_save} seed {config.seed} max_norm_zs {config.max_norm_zs[-1]} noise_shift_delta {config.noise_shift_delta} \
|
139 |
+
scheduler_type {config.scheduler_type} fp16 {args.fp16}\
|
140 |
+
"""
|
my_run.py
ADDED
@@ -0,0 +1,476 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from diffusers import AutoPipelineForImage2Image
|
2 |
+
from diffusers import DDPMScheduler
|
3 |
+
from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img import retrieve_timesteps, retrieve_latents
|
4 |
+
from diffusers.schedulers.scheduling_ddim import DDIMSchedulerOutput
|
5 |
+
import torch
|
6 |
+
from PIL import Image
|
7 |
+
|
8 |
+
num_steps_inversion = 5
|
9 |
+
strngth = 0.8
|
10 |
+
generator = None
|
11 |
+
device = "cuda"
|
12 |
+
image_path = "edit_dataset/01.jpg"
|
13 |
+
src_prompt = "butterfly perched on purple flower"
|
14 |
+
tgt_prompt = "dragonfly perched on purple flower"
|
15 |
+
ws1 = [1.5, 1.5, 1.5, 1.5]
|
16 |
+
ws2 = [1, 1, 1, 1]
|
17 |
+
|
18 |
+
def encode_image(image, pipe):
|
19 |
+
image = pipe.image_processor.preprocess(image)
|
20 |
+
image = image.to(device=device, dtype=pipeline.dtype)
|
21 |
+
|
22 |
+
if pipe.vae.config.force_upcast:
|
23 |
+
image = image.float()
|
24 |
+
pipe.vae.to(dtype=torch.float32)
|
25 |
+
|
26 |
+
if isinstance(generator, list):
|
27 |
+
init_latents = [
|
28 |
+
retrieve_latents(pipe.vae.encode(image[i : i + 1]), generator=generator[i])
|
29 |
+
for i in range(1)
|
30 |
+
]
|
31 |
+
init_latents = torch.cat(init_latents, dim=0)
|
32 |
+
else:
|
33 |
+
init_latents = retrieve_latents(pipe.vae.encode(image), generator=generator)
|
34 |
+
|
35 |
+
if pipe.vae.config.force_upcast:
|
36 |
+
pipe.vae.to(pipeline.dtype)
|
37 |
+
|
38 |
+
init_latents = init_latents.to(pipeline.dtype)
|
39 |
+
init_latents = pipe.vae.config.scaling_factor * init_latents
|
40 |
+
|
41 |
+
return init_latents.to(dtype=torch.float16)
|
42 |
+
|
43 |
+
# def create_xts(scheduler, timesteps, x_0, noise_shift_delta=1, generator=None):
|
44 |
+
# noising_delta = noise_shift_delta * (timesteps[0] - timesteps[1])
|
45 |
+
# noise_timesteps = [timestep - int(noising_delta) for timestep in timesteps]
|
46 |
+
# noise_timesteps = noise_timesteps[:3]
|
47 |
+
|
48 |
+
# x_0_expanded = x_0.expand(len(noise_timesteps), -1, -1, -1)
|
49 |
+
# noise = torch.randn(x_0_expanded.size(), generator=generator, device="cpu", dtype=x_0.dtype).to(x_0.device)
|
50 |
+
# x_ts = scheduler.add_noise(x_0_expanded, noise, torch.IntTensor(noise_timesteps))
|
51 |
+
# x_ts = [t.unsqueeze(dim=0) for t in list(x_ts)]
|
52 |
+
# x_ts += [x_0]
|
53 |
+
# return x_ts
|
54 |
+
|
55 |
+
def deterministic_ddpm_step(
|
56 |
+
model_output: torch.FloatTensor,
|
57 |
+
timestep,
|
58 |
+
sample: torch.FloatTensor,
|
59 |
+
eta,
|
60 |
+
use_clipped_model_output,
|
61 |
+
generator,
|
62 |
+
variance_noise,
|
63 |
+
return_dict,
|
64 |
+
scheduler,
|
65 |
+
):
|
66 |
+
"""
|
67 |
+
Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
|
68 |
+
process from the learned model outputs (most often the predicted noise).
|
69 |
+
|
70 |
+
Args:
|
71 |
+
model_output (`torch.FloatTensor`):
|
72 |
+
The direct output from learned diffusion model.
|
73 |
+
timestep (`float`):
|
74 |
+
The current discrete timestep in the diffusion chain.
|
75 |
+
sample (`torch.FloatTensor`):
|
76 |
+
A current instance of a sample created by the diffusion process.
|
77 |
+
generator (`torch.Generator`, *optional*):
|
78 |
+
A random number generator.
|
79 |
+
return_dict (`bool`, *optional*, defaults to `True`):
|
80 |
+
Whether or not to return a [`~schedulers.scheduling_ddpm.DDPMSchedulerOutput`] or `tuple`.
|
81 |
+
|
82 |
+
Returns:
|
83 |
+
[`~schedulers.scheduling_ddpm.DDPMSchedulerOutput`] or `tuple`:
|
84 |
+
If return_dict is `True`, [`~schedulers.scheduling_ddpm.DDPMSchedulerOutput`] is returned, otherwise a
|
85 |
+
tuple is returned where the first element is the sample tensor.
|
86 |
+
|
87 |
+
"""
|
88 |
+
t = timestep
|
89 |
+
|
90 |
+
prev_t = scheduler.previous_timestep(t)
|
91 |
+
|
92 |
+
if model_output.shape[1] == sample.shape[1] * 2 and scheduler.variance_type in [
|
93 |
+
"learned",
|
94 |
+
"learned_range",
|
95 |
+
]:
|
96 |
+
model_output, predicted_variance = torch.split(
|
97 |
+
model_output, sample.shape[1], dim=1
|
98 |
+
)
|
99 |
+
else:
|
100 |
+
predicted_variance = None
|
101 |
+
|
102 |
+
# 1. compute alphas, betas
|
103 |
+
alpha_prod_t = scheduler.alphas_cumprod[t]
|
104 |
+
alpha_prod_t_prev = (
|
105 |
+
scheduler.alphas_cumprod[prev_t] if prev_t >= 0 else scheduler.one
|
106 |
+
)
|
107 |
+
beta_prod_t = 1 - alpha_prod_t
|
108 |
+
beta_prod_t_prev = 1 - alpha_prod_t_prev
|
109 |
+
current_alpha_t = alpha_prod_t / alpha_prod_t_prev
|
110 |
+
current_beta_t = 1 - current_alpha_t
|
111 |
+
|
112 |
+
# 2. compute predicted original sample from predicted noise also called
|
113 |
+
# "predicted x_0" of formula (15) from https://arxiv.org/pdf/2006.11239.pdf
|
114 |
+
if scheduler.config.prediction_type == "epsilon":
|
115 |
+
pred_original_sample = (
|
116 |
+
sample - beta_prod_t ** (0.5) * model_output
|
117 |
+
) / alpha_prod_t ** (0.5)
|
118 |
+
elif scheduler.config.prediction_type == "sample":
|
119 |
+
pred_original_sample = model_output
|
120 |
+
elif scheduler.config.prediction_type == "v_prediction":
|
121 |
+
pred_original_sample = (alpha_prod_t**0.5) * sample - (
|
122 |
+
beta_prod_t**0.5
|
123 |
+
) * model_output
|
124 |
+
else:
|
125 |
+
raise ValueError(
|
126 |
+
f"prediction_type given as {scheduler.config.prediction_type} must be one of `epsilon`, `sample` or"
|
127 |
+
" `v_prediction` for the DDPMScheduler."
|
128 |
+
)
|
129 |
+
|
130 |
+
# 3. Clip or threshold "predicted x_0"
|
131 |
+
if scheduler.config.thresholding:
|
132 |
+
pred_original_sample = scheduler._threshold_sample(pred_original_sample)
|
133 |
+
elif scheduler.config.clip_sample:
|
134 |
+
pred_original_sample = pred_original_sample.clamp(
|
135 |
+
-scheduler.config.clip_sample_range, scheduler.config.clip_sample_range
|
136 |
+
)
|
137 |
+
|
138 |
+
# 4. Compute coefficients for pred_original_sample x_0 and current sample x_t
|
139 |
+
# See formula (7) from https://arxiv.org/pdf/2006.11239.pdf
|
140 |
+
pred_original_sample_coeff = (
|
141 |
+
alpha_prod_t_prev ** (0.5) * current_beta_t
|
142 |
+
) / beta_prod_t
|
143 |
+
current_sample_coeff = current_alpha_t ** (0.5) * beta_prod_t_prev / beta_prod_t
|
144 |
+
|
145 |
+
# 5. Compute predicted previous sample ยต_t
|
146 |
+
# See formula (7) from https://arxiv.org/pdf/2006.11239.pdf
|
147 |
+
pred_prev_sample = (
|
148 |
+
pred_original_sample_coeff * pred_original_sample
|
149 |
+
+ current_sample_coeff * sample
|
150 |
+
)
|
151 |
+
|
152 |
+
return pred_prev_sample
|
153 |
+
|
154 |
+
def normalize(
|
155 |
+
z_t,
|
156 |
+
i,
|
157 |
+
max_norm_zs,
|
158 |
+
):
|
159 |
+
max_norm = max_norm_zs[i]
|
160 |
+
if max_norm < 0:
|
161 |
+
return z_t, 1
|
162 |
+
|
163 |
+
norm = torch.norm(z_t)
|
164 |
+
if norm < max_norm:
|
165 |
+
return z_t, 1
|
166 |
+
|
167 |
+
coeff = max_norm / norm
|
168 |
+
z_t = z_t * coeff
|
169 |
+
return z_t, coeff
|
170 |
+
|
171 |
+
def step_save_latents(
|
172 |
+
self,
|
173 |
+
model_output: torch.FloatTensor,
|
174 |
+
timestep: int,
|
175 |
+
sample: torch.FloatTensor,
|
176 |
+
eta: float = 0.0,
|
177 |
+
use_clipped_model_output: bool = False,
|
178 |
+
generator=None,
|
179 |
+
variance_noise= None,
|
180 |
+
return_dict: bool = True,
|
181 |
+
):
|
182 |
+
|
183 |
+
timestep_index = self._inner_index
|
184 |
+
next_timestep_index = timestep_index + 1
|
185 |
+
u_hat_t = deterministic_ddpm_step(
|
186 |
+
model_output=model_output,
|
187 |
+
timestep=timestep,
|
188 |
+
sample=sample,
|
189 |
+
eta=eta,
|
190 |
+
use_clipped_model_output=use_clipped_model_output,
|
191 |
+
generator=generator,
|
192 |
+
variance_noise=variance_noise,
|
193 |
+
return_dict=False,
|
194 |
+
scheduler=self,
|
195 |
+
)
|
196 |
+
x_t_minus_1 = self.x_ts[timestep_index]
|
197 |
+
self.x_ts_c_hat.append(u_hat_t)
|
198 |
+
|
199 |
+
z_t = x_t_minus_1 - u_hat_t
|
200 |
+
self.latents.append(z_t)
|
201 |
+
|
202 |
+
z_t, _ = normalize(z_t, timestep_index, [-1, -1, -1, 15.5])
|
203 |
+
x_t_minus_1_predicted = u_hat_t + z_t
|
204 |
+
|
205 |
+
if not return_dict:
|
206 |
+
return (x_t_minus_1_predicted,)
|
207 |
+
|
208 |
+
return DDIMSchedulerOutput(prev_sample=x_t_minus_1, pred_original_sample=None)
|
209 |
+
|
210 |
+
def step_use_latents(
|
211 |
+
self,
|
212 |
+
model_output: torch.FloatTensor,
|
213 |
+
timestep: int,
|
214 |
+
sample: torch.FloatTensor,
|
215 |
+
eta: float = 0.0,
|
216 |
+
use_clipped_model_output: bool = False,
|
217 |
+
generator=None,
|
218 |
+
variance_noise= None,
|
219 |
+
return_dict: bool = True,
|
220 |
+
):
|
221 |
+
print(f'_inner_index: {self._inner_index}')
|
222 |
+
timestep_index = self._inner_index
|
223 |
+
next_timestep_index = timestep_index + 1
|
224 |
+
z_t = self.latents[timestep_index] # + 1 because latents[0] is X_T
|
225 |
+
|
226 |
+
_, normalize_coefficient = normalize(
|
227 |
+
z_t,
|
228 |
+
timestep_index,
|
229 |
+
[-1, -1, -1, 15.5],
|
230 |
+
)
|
231 |
+
|
232 |
+
if normalize_coefficient == 0:
|
233 |
+
eta = 0
|
234 |
+
|
235 |
+
# eta = normalize_coefficient
|
236 |
+
|
237 |
+
x_t_hat_c_hat = deterministic_ddpm_step(
|
238 |
+
model_output=model_output,
|
239 |
+
timestep=timestep,
|
240 |
+
sample=sample,
|
241 |
+
eta=eta,
|
242 |
+
use_clipped_model_output=use_clipped_model_output,
|
243 |
+
generator=generator,
|
244 |
+
variance_noise=variance_noise,
|
245 |
+
return_dict=False,
|
246 |
+
scheduler=self,
|
247 |
+
)
|
248 |
+
|
249 |
+
w1 = ws1[timestep_index]
|
250 |
+
w2 = ws2[timestep_index]
|
251 |
+
|
252 |
+
x_t_minus_1_exact = self.x_ts[timestep_index]
|
253 |
+
x_t_minus_1_exact = x_t_minus_1_exact.expand_as(x_t_hat_c_hat)
|
254 |
+
|
255 |
+
x_t_c_hat: torch.Tensor = self.x_ts_c_hat[timestep_index]
|
256 |
+
|
257 |
+
x_t_c = x_t_c_hat[0].expand_as(x_t_hat_c_hat)
|
258 |
+
|
259 |
+
zero_index_reconstruction = 0
|
260 |
+
edit_prompts_num = (model_output.size(0) - zero_index_reconstruction) // 2
|
261 |
+
x_t_hat_c_indices = (zero_index_reconstruction, edit_prompts_num + zero_index_reconstruction)
|
262 |
+
edit_images_indices = (
|
263 |
+
edit_prompts_num + zero_index_reconstruction,
|
264 |
+
model_output.size(0)
|
265 |
+
)
|
266 |
+
x_t_hat_c = torch.zeros_like(x_t_hat_c_hat)
|
267 |
+
x_t_hat_c[edit_images_indices[0] : edit_images_indices[1]] = x_t_hat_c_hat[
|
268 |
+
x_t_hat_c_indices[0] : x_t_hat_c_indices[1]
|
269 |
+
]
|
270 |
+
v1 = x_t_hat_c_hat - x_t_hat_c
|
271 |
+
v2 = x_t_hat_c - normalize_coefficient * x_t_c
|
272 |
+
|
273 |
+
x_t_minus_1 = normalize_coefficient * x_t_minus_1_exact + w1 * v1 + w2 * v2
|
274 |
+
|
275 |
+
x_t_minus_1[x_t_hat_c_indices[0] : x_t_hat_c_indices[1]] = x_t_minus_1[
|
276 |
+
edit_images_indices[0] : edit_images_indices[1]
|
277 |
+
] # update x_t_hat_c to be x_t_hat_c_hat
|
278 |
+
|
279 |
+
|
280 |
+
if not return_dict:
|
281 |
+
return (x_t_minus_1,)
|
282 |
+
|
283 |
+
return DDIMSchedulerOutput(
|
284 |
+
prev_sample=x_t_minus_1,
|
285 |
+
pred_original_sample=None,
|
286 |
+
)
|
287 |
+
|
288 |
+
|
289 |
+
class myDDPMScheduler(DDPMScheduler):
|
290 |
+
def step(
|
291 |
+
self,
|
292 |
+
model_output: torch.FloatTensor,
|
293 |
+
timestep: int,
|
294 |
+
sample: torch.FloatTensor,
|
295 |
+
eta: float = 0.0,
|
296 |
+
use_clipped_model_output: bool = False,
|
297 |
+
generator=None,
|
298 |
+
variance_noise= None,
|
299 |
+
return_dict: bool = True,
|
300 |
+
):
|
301 |
+
print(f"timestep: {timestep}")
|
302 |
+
|
303 |
+
res_inv = step_save_latents(
|
304 |
+
self,
|
305 |
+
model_output[:1, :, :, :],
|
306 |
+
timestep,
|
307 |
+
sample[:1, :, :, :],
|
308 |
+
eta,
|
309 |
+
use_clipped_model_output,
|
310 |
+
generator,
|
311 |
+
variance_noise,
|
312 |
+
return_dict,
|
313 |
+
)
|
314 |
+
|
315 |
+
res_inf = step_use_latents(
|
316 |
+
self,
|
317 |
+
model_output[1:, :, :, :],
|
318 |
+
timestep,
|
319 |
+
sample[1:, :, :, :],
|
320 |
+
eta,
|
321 |
+
use_clipped_model_output,
|
322 |
+
generator,
|
323 |
+
variance_noise,
|
324 |
+
return_dict,
|
325 |
+
)
|
326 |
+
|
327 |
+
self._inner_index+=1
|
328 |
+
|
329 |
+
res = (torch.cat((res_inv[0], res_inf[0]), dim=0),)
|
330 |
+
return res
|
331 |
+
|
332 |
+
|
333 |
+
pipeline = AutoPipelineForImage2Image.from_pretrained("stabilityai/sdxl-turbo", torch_dtype=torch.float16, variant="fp16", safety_checker = None)
|
334 |
+
pipeline = pipeline.to("cuda")
|
335 |
+
pipeline.scheduler = DDPMScheduler.from_pretrained( # type: ignore
|
336 |
+
'stabilityai/sdxl-turbo',
|
337 |
+
subfolder="scheduler",
|
338 |
+
# cache_dir="/home/joberant/NLP_2223/giladd/test_dir/sdxl-turbo/models_cache",
|
339 |
+
)
|
340 |
+
# pipeline.scheduler = DDPMScheduler.from_config(pipeline.scheduler.config)
|
341 |
+
|
342 |
+
denoising_start = 0.2
|
343 |
+
timesteps, num_inference_steps = retrieve_timesteps(
|
344 |
+
pipeline.scheduler, num_steps_inversion, device, None
|
345 |
+
)
|
346 |
+
timesteps, num_inference_steps = pipeline.get_timesteps(
|
347 |
+
num_inference_steps=num_inference_steps,
|
348 |
+
device=device,
|
349 |
+
denoising_start=denoising_start,
|
350 |
+
strength=0,
|
351 |
+
)
|
352 |
+
timesteps = timesteps.type(torch.int64)
|
353 |
+
from functools import partial
|
354 |
+
|
355 |
+
timesteps = [torch.tensor(t) for t in timesteps.tolist()]
|
356 |
+
pipeline.__call__ = partial(
|
357 |
+
pipeline.__call__,
|
358 |
+
num_inference_steps=num_steps_inversion,
|
359 |
+
guidance_scale=0,
|
360 |
+
generator=generator,
|
361 |
+
denoising_start=denoising_start,
|
362 |
+
strength=0,
|
363 |
+
)
|
364 |
+
|
365 |
+
# timesteps, num_inference_steps = retrieve_timesteps(pipeline.scheduler, num_steps_inversion, device, None)
|
366 |
+
# timesteps, num_inference_steps = pipeline.get_timesteps(num_inference_steps=num_inference_steps, device=device, strength=strngth)
|
367 |
+
|
368 |
+
|
369 |
+
from utils import get_ddpm_inversion_scheduler, create_xts
|
370 |
+
|
371 |
+
|
372 |
+
|
373 |
+
from config import get_config, get_config_name
|
374 |
+
import argparse
|
375 |
+
|
376 |
+
# parser = argparse.ArgumentParser()
|
377 |
+
# parser.add_argument("--images_paths", type=str, default=None)
|
378 |
+
# parser.add_argument("--images_folder", type=str, default=None)
|
379 |
+
# parser.set_defaults(force_use_cpu=False)
|
380 |
+
# parser.add_argument("--force_use_cpu", action="store_true")
|
381 |
+
# parser.add_argument("--folder_name", type=str, default='test_measure_time')
|
382 |
+
# parser.add_argument("--config_from_file", type=str, default='run_configs/noise_shift_guidance_1_5.yaml')
|
383 |
+
# parser.set_defaults(save_intermediate_results=False)
|
384 |
+
# parser.add_argument("--save_intermediate_results", action="store_true")
|
385 |
+
# parser.add_argument("--batch_size", type=int, default=None)
|
386 |
+
# parser.set_defaults(skip_p_to_p=False)
|
387 |
+
# parser.add_argument("--skip_p_to_p", action="store_true", default=True)
|
388 |
+
# parser.set_defaults(only_p_to_p=False)
|
389 |
+
# parser.add_argument("--only_p_to_p", action="store_true")
|
390 |
+
# parser.set_defaults(fp16=False)
|
391 |
+
# parser.add_argument("--fp16", action="store_true", default=False)
|
392 |
+
# parser.add_argument("--prompts_file", type=str, default='dataset_measure_time/dataset.json')
|
393 |
+
# parser.add_argument("--images_in_prompts_file", type=str, default=None)
|
394 |
+
# parser.add_argument("--seed", type=int, default=2)
|
395 |
+
# parser.add_argument("--time_measure_n", type=int, default=1)
|
396 |
+
|
397 |
+
# args = parser.parse_args()
|
398 |
+
class Object(object):
|
399 |
+
pass
|
400 |
+
|
401 |
+
args = Object()
|
402 |
+
args.images_paths = None
|
403 |
+
args.images_folder = None
|
404 |
+
args.force_use_cpu = False
|
405 |
+
args.folder_name = 'test_measure_time'
|
406 |
+
args.config_from_file = 'run_configs/noise_shift_guidance_1_5.yaml'
|
407 |
+
args.save_intermediate_results = False
|
408 |
+
args.batch_size = None
|
409 |
+
args.skip_p_to_p = True
|
410 |
+
args.only_p_to_p = False
|
411 |
+
args.fp16 = False
|
412 |
+
args.prompts_file = 'dataset_measure_time/dataset.json'
|
413 |
+
args.images_in_prompts_file = None
|
414 |
+
args.seed = 986
|
415 |
+
args.time_measure_n = 1
|
416 |
+
|
417 |
+
|
418 |
+
assert (
|
419 |
+
args.batch_size is None or args.save_intermediate_results is False
|
420 |
+
), "save_intermediate_results is not implemented for batch_size > 1"
|
421 |
+
|
422 |
+
config = get_config(args)
|
423 |
+
|
424 |
+
|
425 |
+
|
426 |
+
|
427 |
+
|
428 |
+
# latent = latents[0].expand(3, -1, -1, -1)
|
429 |
+
# prompt = [src_prompt, src_prompt, tgt_prompt]
|
430 |
+
|
431 |
+
# image = pipeline.__call__(image=latent, prompt=prompt, eta=1).images
|
432 |
+
|
433 |
+
# for i, im in enumerate(image):
|
434 |
+
# im.save(f"output_{i}.png")
|
435 |
+
|
436 |
+
def run(image_path, src_prompt, tgt_prompt, seed, w1, w2):
|
437 |
+
generator = torch.Generator().manual_seed(seed)
|
438 |
+
x_0_image = Image.open(image_path).convert("RGB").resize((512, 512), Image.LANCZOS)
|
439 |
+
x_0 = encode_image(x_0_image, pipeline)
|
440 |
+
# x_ts = create_xts(pipeline.scheduler, timesteps, x_0, noise_shift_delta=1, generator=generator)
|
441 |
+
x_ts = create_xts(1, None, 0, generator, pipeline.scheduler, timesteps, x_0, no_add_noise=False)
|
442 |
+
x_ts = [xt.to(dtype=torch.float16) for xt in x_ts]
|
443 |
+
latents = [x_ts[0]]
|
444 |
+
x_ts_c_hat = [None]
|
445 |
+
config.ws1 = [w1] * 4
|
446 |
+
config.ws2 = [w2] * 4
|
447 |
+
pipeline.scheduler = get_ddpm_inversion_scheduler(
|
448 |
+
pipeline.scheduler,
|
449 |
+
config.step_function,
|
450 |
+
config,
|
451 |
+
timesteps,
|
452 |
+
config.save_timesteps,
|
453 |
+
latents,
|
454 |
+
x_ts,
|
455 |
+
x_ts_c_hat,
|
456 |
+
args.save_intermediate_results,
|
457 |
+
pipeline,
|
458 |
+
x_0,
|
459 |
+
v1s_images := [],
|
460 |
+
v2s_images := [],
|
461 |
+
deltas_images := [],
|
462 |
+
v1_x0s := [],
|
463 |
+
v2_x0s := [],
|
464 |
+
deltas_x0s := [],
|
465 |
+
"res12",
|
466 |
+
image_name="im_name",
|
467 |
+
time_measure_n=args.time_measure_n,
|
468 |
+
)
|
469 |
+
latent = latents[0].expand(3, -1, -1, -1)
|
470 |
+
prompt = [src_prompt, src_prompt, tgt_prompt]
|
471 |
+
image = pipeline.__call__(image=latent, prompt=prompt, eta=1).images
|
472 |
+
return image[2]
|
473 |
+
|
474 |
+
if __name__ == "__main__":
|
475 |
+
res = run(image_path, src_prompt, tgt_prompt, args.seed, 1.5, 1.0)
|
476 |
+
res.save("output.png")
|
resize.py
ADDED
@@ -0,0 +1,18 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from PIL import Image
|
2 |
+
|
3 |
+
def resize_image(input_path, output_path, new_size):
|
4 |
+
# Open the image
|
5 |
+
image = Image.open(input_path)
|
6 |
+
|
7 |
+
# Resize the image
|
8 |
+
resized_image = image.resize(new_size)
|
9 |
+
|
10 |
+
# Save the resized image
|
11 |
+
resized_image.save(output_path)
|
12 |
+
|
13 |
+
# Example usage
|
14 |
+
input_path = "demo_im/pexels-pixabay-458976.png"
|
15 |
+
output_path = "demo_im/pexels-pixabay-458976.less.png"
|
16 |
+
new_size = (512, 512)
|
17 |
+
|
18 |
+
resize_image(input_path, output_path, new_size)
|
utils.py
ADDED
@@ -0,0 +1,1356 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import itertools
|
2 |
+
from typing import List, Optional, Union
|
3 |
+
import PIL
|
4 |
+
import PIL.Image
|
5 |
+
import torch
|
6 |
+
from diffusers.schedulers.scheduling_ddim import DDIMSchedulerOutput
|
7 |
+
from diffusers.utils import make_image_grid
|
8 |
+
from PIL import Image, ImageDraw, ImageFont
|
9 |
+
import os
|
10 |
+
from diffusers.utils import (
|
11 |
+
logging,
|
12 |
+
USE_PEFT_BACKEND,
|
13 |
+
scale_lora_layers,
|
14 |
+
unscale_lora_layers,
|
15 |
+
)
|
16 |
+
from diffusers.loaders import (
|
17 |
+
StableDiffusionXLLoraLoaderMixin,
|
18 |
+
)
|
19 |
+
from diffusers.image_processor import VaeImageProcessor
|
20 |
+
|
21 |
+
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
|
22 |
+
|
23 |
+
from diffusers.models.lora import adjust_lora_scale_text_encoder
|
24 |
+
from diffusers import DiffusionPipeline
|
25 |
+
|
26 |
+
|
27 |
+
VECTOR_DATA_FOLDER = "vector_data"
|
28 |
+
VECTOR_DATA_DICT = "vector_data"
|
29 |
+
|
30 |
+
|
31 |
+
def encode_image(image: PIL.Image, pipe: DiffusionPipeline):
|
32 |
+
pipe.image_processor: VaeImageProcessor = pipe.image_processor # type: ignore
|
33 |
+
image = pipe.image_processor.pil_to_numpy(image)
|
34 |
+
image = pipe.image_processor.numpy_to_pt(image)
|
35 |
+
image = image.to(pipe.device)
|
36 |
+
return (
|
37 |
+
pipe.vae.encode(
|
38 |
+
pipe.image_processor.preprocess(image),
|
39 |
+
).latent_dist.mode()
|
40 |
+
* pipe.vae.config.scaling_factor
|
41 |
+
)
|
42 |
+
|
43 |
+
|
44 |
+
def decode_latents(latent, pipe):
|
45 |
+
latent_img = pipe.vae.decode(
|
46 |
+
latent / pipe.vae.config.scaling_factor, return_dict=False
|
47 |
+
)[0]
|
48 |
+
return pipe.image_processor.postprocess(latent_img, output_type="pil")
|
49 |
+
|
50 |
+
|
51 |
+
def get_device(argv, args=None):
|
52 |
+
import sys
|
53 |
+
|
54 |
+
def debugger_is_active():
|
55 |
+
return hasattr(sys, "gettrace") and sys.gettrace() is not None
|
56 |
+
|
57 |
+
if args:
|
58 |
+
return (
|
59 |
+
torch.device("cuda")
|
60 |
+
if (torch.cuda.is_available() and not debugger_is_active())
|
61 |
+
and not args.force_use_cpu
|
62 |
+
else torch.device("cpu")
|
63 |
+
)
|
64 |
+
|
65 |
+
return (
|
66 |
+
torch.device("cuda")
|
67 |
+
if (torch.cuda.is_available() and not debugger_is_active())
|
68 |
+
and not "cpu" in set(argv[1:])
|
69 |
+
else torch.device("cpu")
|
70 |
+
)
|
71 |
+
|
72 |
+
|
73 |
+
def deterministic_ddim_step(
|
74 |
+
model_output: torch.FloatTensor,
|
75 |
+
timestep: int,
|
76 |
+
sample: torch.FloatTensor,
|
77 |
+
eta: float = 0.0,
|
78 |
+
use_clipped_model_output: bool = False,
|
79 |
+
generator=None,
|
80 |
+
variance_noise: Optional[torch.FloatTensor] = None,
|
81 |
+
return_dict: bool = True,
|
82 |
+
scheduler=None,
|
83 |
+
):
|
84 |
+
|
85 |
+
if scheduler.num_inference_steps is None:
|
86 |
+
raise ValueError(
|
87 |
+
"Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
|
88 |
+
)
|
89 |
+
|
90 |
+
prev_timestep = (
|
91 |
+
timestep - scheduler.config.num_train_timesteps // scheduler.num_inference_steps
|
92 |
+
)
|
93 |
+
|
94 |
+
# 2. compute alphas, betas
|
95 |
+
alpha_prod_t = scheduler.alphas_cumprod[timestep]
|
96 |
+
alpha_prod_t_prev = (
|
97 |
+
scheduler.alphas_cumprod[prev_timestep]
|
98 |
+
if prev_timestep >= 0
|
99 |
+
else scheduler.final_alpha_cumprod
|
100 |
+
)
|
101 |
+
|
102 |
+
beta_prod_t = 1 - alpha_prod_t
|
103 |
+
|
104 |
+
if scheduler.config.prediction_type == "epsilon":
|
105 |
+
pred_original_sample = (
|
106 |
+
sample - beta_prod_t ** (0.5) * model_output
|
107 |
+
) / alpha_prod_t ** (0.5)
|
108 |
+
pred_epsilon = model_output
|
109 |
+
elif scheduler.config.prediction_type == "sample":
|
110 |
+
pred_original_sample = model_output
|
111 |
+
pred_epsilon = (
|
112 |
+
sample - alpha_prod_t ** (0.5) * pred_original_sample
|
113 |
+
) / beta_prod_t ** (0.5)
|
114 |
+
elif scheduler.config.prediction_type == "v_prediction":
|
115 |
+
pred_original_sample = (alpha_prod_t**0.5) * sample - (
|
116 |
+
beta_prod_t**0.5
|
117 |
+
) * model_output
|
118 |
+
pred_epsilon = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample
|
119 |
+
else:
|
120 |
+
raise ValueError(
|
121 |
+
f"prediction_type given as {scheduler.config.prediction_type} must be one of `epsilon`, `sample`, or"
|
122 |
+
" `v_prediction`"
|
123 |
+
)
|
124 |
+
|
125 |
+
# 4. Clip or threshold "predicted x_0"
|
126 |
+
if scheduler.config.thresholding:
|
127 |
+
pred_original_sample = scheduler._threshold_sample(pred_original_sample)
|
128 |
+
elif scheduler.config.clip_sample:
|
129 |
+
pred_original_sample = pred_original_sample.clamp(
|
130 |
+
-scheduler.config.clip_sample_range,
|
131 |
+
scheduler.config.clip_sample_range,
|
132 |
+
)
|
133 |
+
|
134 |
+
# 5. compute variance: "sigma_t(ฮท)" -> see formula (16)
|
135 |
+
# ฯ_t = sqrt((1 โ ฮฑ_tโ1)/(1 โ ฮฑ_t)) * sqrt(1 โ ฮฑ_t/ฮฑ_tโ1)
|
136 |
+
variance = scheduler._get_variance(timestep, prev_timestep)
|
137 |
+
std_dev_t = eta * variance ** (0.5)
|
138 |
+
|
139 |
+
if use_clipped_model_output:
|
140 |
+
# the pred_epsilon is always re-derived from the clipped x_0 in Glide
|
141 |
+
pred_epsilon = (
|
142 |
+
sample - alpha_prod_t ** (0.5) * pred_original_sample
|
143 |
+
) / beta_prod_t ** (0.5)
|
144 |
+
|
145 |
+
# 6. compute "direction pointing to x_t" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
|
146 |
+
pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** (
|
147 |
+
0.5
|
148 |
+
) * pred_epsilon
|
149 |
+
|
150 |
+
# 7. compute x_t without "random noise" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
|
151 |
+
prev_sample = (
|
152 |
+
alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction
|
153 |
+
)
|
154 |
+
return prev_sample
|
155 |
+
|
156 |
+
|
157 |
+
def deterministic_euler_step(
|
158 |
+
model_output: torch.FloatTensor,
|
159 |
+
timestep: Union[float, torch.FloatTensor],
|
160 |
+
sample: torch.FloatTensor,
|
161 |
+
eta,
|
162 |
+
use_clipped_model_output,
|
163 |
+
generator,
|
164 |
+
variance_noise,
|
165 |
+
return_dict,
|
166 |
+
scheduler,
|
167 |
+
):
|
168 |
+
"""
|
169 |
+
Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
|
170 |
+
process from the learned model outputs (most often the predicted noise).
|
171 |
+
|
172 |
+
Args:
|
173 |
+
model_output (`torch.FloatTensor`):
|
174 |
+
The direct output from learned diffusion model.
|
175 |
+
timestep (`float`):
|
176 |
+
The current discrete timestep in the diffusion chain.
|
177 |
+
sample (`torch.FloatTensor`):
|
178 |
+
A current instance of a sample created by the diffusion process.
|
179 |
+
generator (`torch.Generator`, *optional*):
|
180 |
+
A random number generator.
|
181 |
+
return_dict (`bool`):
|
182 |
+
Whether or not to return a
|
183 |
+
[`~schedulers.scheduling_euler_ancestral_discrete.EulerAncestralDiscreteSchedulerOutput`] or tuple.
|
184 |
+
|
185 |
+
Returns:
|
186 |
+
[`~schedulers.scheduling_euler_ancestral_discrete.EulerAncestralDiscreteSchedulerOutput`] or `tuple`:
|
187 |
+
If return_dict is `True`,
|
188 |
+
[`~schedulers.scheduling_euler_ancestral_discrete.EulerAncestralDiscreteSchedulerOutput`] is returned,
|
189 |
+
otherwise a tuple is returned where the first element is the sample tensor.
|
190 |
+
|
191 |
+
"""
|
192 |
+
|
193 |
+
if (
|
194 |
+
isinstance(timestep, int)
|
195 |
+
or isinstance(timestep, torch.IntTensor)
|
196 |
+
or isinstance(timestep, torch.LongTensor)
|
197 |
+
):
|
198 |
+
raise ValueError(
|
199 |
+
(
|
200 |
+
"Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to"
|
201 |
+
" `EulerDiscreteScheduler.step()` is not supported. Make sure to pass"
|
202 |
+
" one of the `scheduler.timesteps` as a timestep."
|
203 |
+
),
|
204 |
+
)
|
205 |
+
|
206 |
+
if scheduler.step_index is None:
|
207 |
+
scheduler._init_step_index(timestep)
|
208 |
+
|
209 |
+
sigma = scheduler.sigmas[scheduler.step_index]
|
210 |
+
|
211 |
+
# Upcast to avoid precision issues when computing prev_sample
|
212 |
+
sample = sample.to(torch.float32)
|
213 |
+
|
214 |
+
# 1. compute predicted original sample (x_0) from sigma-scaled predicted noise
|
215 |
+
if scheduler.config.prediction_type == "epsilon":
|
216 |
+
pred_original_sample = sample - sigma * model_output
|
217 |
+
elif scheduler.config.prediction_type == "v_prediction":
|
218 |
+
# * c_out + input * c_skip
|
219 |
+
pred_original_sample = model_output * (-sigma / (sigma**2 + 1) ** 0.5) + (
|
220 |
+
sample / (sigma**2 + 1)
|
221 |
+
)
|
222 |
+
elif scheduler.config.prediction_type == "sample":
|
223 |
+
raise NotImplementedError("prediction_type not implemented yet: sample")
|
224 |
+
else:
|
225 |
+
raise ValueError(
|
226 |
+
f"prediction_type given as {scheduler.config.prediction_type} must be one of `epsilon`, or `v_prediction`"
|
227 |
+
)
|
228 |
+
|
229 |
+
sigma_from = scheduler.sigmas[scheduler.step_index]
|
230 |
+
sigma_to = scheduler.sigmas[scheduler.step_index + 1]
|
231 |
+
sigma_up = (sigma_to**2 * (sigma_from**2 - sigma_to**2) / sigma_from**2) ** 0.5
|
232 |
+
sigma_down = (sigma_to**2 - sigma_up**2) ** 0.5
|
233 |
+
|
234 |
+
# 2. Convert to an ODE derivative
|
235 |
+
derivative = (sample - pred_original_sample) / sigma
|
236 |
+
|
237 |
+
dt = sigma_down - sigma
|
238 |
+
|
239 |
+
prev_sample = sample + derivative * dt
|
240 |
+
|
241 |
+
# Cast sample back to model compatible dtype
|
242 |
+
prev_sample = prev_sample.to(model_output.dtype)
|
243 |
+
|
244 |
+
# upon completion increase step index by one
|
245 |
+
scheduler._step_index += 1
|
246 |
+
|
247 |
+
return prev_sample
|
248 |
+
|
249 |
+
|
250 |
+
def deterministic_non_ancestral_euler_step(
|
251 |
+
model_output: torch.FloatTensor,
|
252 |
+
timestep: Union[float, torch.FloatTensor],
|
253 |
+
sample: torch.FloatTensor,
|
254 |
+
eta: float = 0.0,
|
255 |
+
use_clipped_model_output: bool = False,
|
256 |
+
s_churn: float = 0.0,
|
257 |
+
s_tmin: float = 0.0,
|
258 |
+
s_tmax: float = float("inf"),
|
259 |
+
s_noise: float = 1.0,
|
260 |
+
generator: Optional[torch.Generator] = None,
|
261 |
+
variance_noise: Optional[torch.FloatTensor] = None,
|
262 |
+
return_dict: bool = True,
|
263 |
+
scheduler=None,
|
264 |
+
):
|
265 |
+
"""
|
266 |
+
Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
|
267 |
+
process from the learned model outputs (most often the predicted noise).
|
268 |
+
|
269 |
+
Args:
|
270 |
+
model_output (`torch.FloatTensor`):
|
271 |
+
The direct output from learned diffusion model.
|
272 |
+
timestep (`float`):
|
273 |
+
The current discrete timestep in the diffusion chain.
|
274 |
+
sample (`torch.FloatTensor`):
|
275 |
+
A current instance of a sample created by the diffusion process.
|
276 |
+
s_churn (`float`):
|
277 |
+
s_tmin (`float`):
|
278 |
+
s_tmax (`float`):
|
279 |
+
s_noise (`float`, defaults to 1.0):
|
280 |
+
Scaling factor for noise added to the sample.
|
281 |
+
generator (`torch.Generator`, *optional*):
|
282 |
+
A random number generator.
|
283 |
+
return_dict (`bool`):
|
284 |
+
Whether or not to return a [`~schedulers.scheduling_euler_discrete.EulerDiscreteSchedulerOutput`] or
|
285 |
+
tuple.
|
286 |
+
|
287 |
+
Returns:
|
288 |
+
[`~schedulers.scheduling_euler_discrete.EulerDiscreteSchedulerOutput`] or `tuple`:
|
289 |
+
If return_dict is `True`, [`~schedulers.scheduling_euler_discrete.EulerDiscreteSchedulerOutput`] is
|
290 |
+
returned, otherwise a tuple is returned where the first element is the sample tensor.
|
291 |
+
"""
|
292 |
+
|
293 |
+
if (
|
294 |
+
isinstance(timestep, int)
|
295 |
+
or isinstance(timestep, torch.IntTensor)
|
296 |
+
or isinstance(timestep, torch.LongTensor)
|
297 |
+
):
|
298 |
+
raise ValueError(
|
299 |
+
(
|
300 |
+
"Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to"
|
301 |
+
" `EulerDiscreteScheduler.step()` is not supported. Make sure to pass"
|
302 |
+
" one of the `scheduler.timesteps` as a timestep."
|
303 |
+
),
|
304 |
+
)
|
305 |
+
|
306 |
+
if not scheduler.is_scale_input_called:
|
307 |
+
logger.warning(
|
308 |
+
"The `scale_model_input` function should be called before `step` to ensure correct denoising. "
|
309 |
+
"See `StableDiffusionPipeline` for a usage example."
|
310 |
+
)
|
311 |
+
|
312 |
+
if scheduler.step_index is None:
|
313 |
+
scheduler._init_step_index(timestep)
|
314 |
+
|
315 |
+
# Upcast to avoid precision issues when computing prev_sample
|
316 |
+
sample = sample.to(torch.float32)
|
317 |
+
|
318 |
+
sigma = scheduler.sigmas[scheduler.step_index]
|
319 |
+
|
320 |
+
gamma = (
|
321 |
+
min(s_churn / (len(scheduler.sigmas) - 1), 2**0.5 - 1)
|
322 |
+
if s_tmin <= sigma <= s_tmax
|
323 |
+
else 0.0
|
324 |
+
)
|
325 |
+
|
326 |
+
sigma_hat = sigma * (gamma + 1)
|
327 |
+
|
328 |
+
# 1. compute predicted original sample (x_0) from sigma-scaled predicted noise
|
329 |
+
# NOTE: "original_sample" should not be an expected prediction_type but is left in for
|
330 |
+
# backwards compatibility
|
331 |
+
if (
|
332 |
+
scheduler.config.prediction_type == "original_sample"
|
333 |
+
or scheduler.config.prediction_type == "sample"
|
334 |
+
):
|
335 |
+
pred_original_sample = model_output
|
336 |
+
elif scheduler.config.prediction_type == "epsilon":
|
337 |
+
pred_original_sample = sample - sigma_hat * model_output
|
338 |
+
elif scheduler.config.prediction_type == "v_prediction":
|
339 |
+
# denoised = model_output * c_out + input * c_skip
|
340 |
+
pred_original_sample = model_output * (-sigma / (sigma**2 + 1) ** 0.5) + (
|
341 |
+
sample / (sigma**2 + 1)
|
342 |
+
)
|
343 |
+
else:
|
344 |
+
raise ValueError(
|
345 |
+
f"prediction_type given as {scheduler.config.prediction_type} must be one of `epsilon`, or `v_prediction`"
|
346 |
+
)
|
347 |
+
|
348 |
+
# 2. Convert to an ODE derivative
|
349 |
+
derivative = (sample - pred_original_sample) / sigma_hat
|
350 |
+
|
351 |
+
dt = scheduler.sigmas[scheduler.step_index + 1] - sigma_hat
|
352 |
+
|
353 |
+
prev_sample = sample + derivative * dt
|
354 |
+
|
355 |
+
# Cast sample back to model compatible dtype
|
356 |
+
prev_sample = prev_sample.to(model_output.dtype)
|
357 |
+
|
358 |
+
# upon completion increase step index by one
|
359 |
+
scheduler._step_index += 1
|
360 |
+
|
361 |
+
return prev_sample
|
362 |
+
|
363 |
+
|
364 |
+
def deterministic_ddpm_step(
|
365 |
+
model_output: torch.FloatTensor,
|
366 |
+
timestep: Union[float, torch.FloatTensor],
|
367 |
+
sample: torch.FloatTensor,
|
368 |
+
eta,
|
369 |
+
use_clipped_model_output,
|
370 |
+
generator,
|
371 |
+
variance_noise,
|
372 |
+
return_dict,
|
373 |
+
scheduler,
|
374 |
+
):
|
375 |
+
"""
|
376 |
+
Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
|
377 |
+
process from the learned model outputs (most often the predicted noise).
|
378 |
+
|
379 |
+
Args:
|
380 |
+
model_output (`torch.FloatTensor`):
|
381 |
+
The direct output from learned diffusion model.
|
382 |
+
timestep (`float`):
|
383 |
+
The current discrete timestep in the diffusion chain.
|
384 |
+
sample (`torch.FloatTensor`):
|
385 |
+
A current instance of a sample created by the diffusion process.
|
386 |
+
generator (`torch.Generator`, *optional*):
|
387 |
+
A random number generator.
|
388 |
+
return_dict (`bool`, *optional*, defaults to `True`):
|
389 |
+
Whether or not to return a [`~schedulers.scheduling_ddpm.DDPMSchedulerOutput`] or `tuple`.
|
390 |
+
|
391 |
+
Returns:
|
392 |
+
[`~schedulers.scheduling_ddpm.DDPMSchedulerOutput`] or `tuple`:
|
393 |
+
If return_dict is `True`, [`~schedulers.scheduling_ddpm.DDPMSchedulerOutput`] is returned, otherwise a
|
394 |
+
tuple is returned where the first element is the sample tensor.
|
395 |
+
|
396 |
+
"""
|
397 |
+
t = timestep
|
398 |
+
|
399 |
+
prev_t = scheduler.previous_timestep(t)
|
400 |
+
|
401 |
+
if model_output.shape[1] == sample.shape[1] * 2 and scheduler.variance_type in [
|
402 |
+
"learned",
|
403 |
+
"learned_range",
|
404 |
+
]:
|
405 |
+
model_output, predicted_variance = torch.split(
|
406 |
+
model_output, sample.shape[1], dim=1
|
407 |
+
)
|
408 |
+
else:
|
409 |
+
predicted_variance = None
|
410 |
+
|
411 |
+
# 1. compute alphas, betas
|
412 |
+
alpha_prod_t = scheduler.alphas_cumprod[t]
|
413 |
+
alpha_prod_t_prev = (
|
414 |
+
scheduler.alphas_cumprod[prev_t] if prev_t >= 0 else scheduler.one
|
415 |
+
)
|
416 |
+
beta_prod_t = 1 - alpha_prod_t
|
417 |
+
beta_prod_t_prev = 1 - alpha_prod_t_prev
|
418 |
+
current_alpha_t = alpha_prod_t / alpha_prod_t_prev
|
419 |
+
current_beta_t = 1 - current_alpha_t
|
420 |
+
|
421 |
+
# 2. compute predicted original sample from predicted noise also called
|
422 |
+
# "predicted x_0" of formula (15) from https://arxiv.org/pdf/2006.11239.pdf
|
423 |
+
if scheduler.config.prediction_type == "epsilon":
|
424 |
+
pred_original_sample = (
|
425 |
+
sample - beta_prod_t ** (0.5) * model_output
|
426 |
+
) / alpha_prod_t ** (0.5)
|
427 |
+
elif scheduler.config.prediction_type == "sample":
|
428 |
+
pred_original_sample = model_output
|
429 |
+
elif scheduler.config.prediction_type == "v_prediction":
|
430 |
+
pred_original_sample = (alpha_prod_t**0.5) * sample - (
|
431 |
+
beta_prod_t**0.5
|
432 |
+
) * model_output
|
433 |
+
else:
|
434 |
+
raise ValueError(
|
435 |
+
f"prediction_type given as {scheduler.config.prediction_type} must be one of `epsilon`, `sample` or"
|
436 |
+
" `v_prediction` for the DDPMScheduler."
|
437 |
+
)
|
438 |
+
|
439 |
+
# 3. Clip or threshold "predicted x_0"
|
440 |
+
if scheduler.config.thresholding:
|
441 |
+
pred_original_sample = scheduler._threshold_sample(pred_original_sample)
|
442 |
+
elif scheduler.config.clip_sample:
|
443 |
+
pred_original_sample = pred_original_sample.clamp(
|
444 |
+
-scheduler.config.clip_sample_range, scheduler.config.clip_sample_range
|
445 |
+
)
|
446 |
+
|
447 |
+
# 4. Compute coefficients for pred_original_sample x_0 and current sample x_t
|
448 |
+
# See formula (7) from https://arxiv.org/pdf/2006.11239.pdf
|
449 |
+
pred_original_sample_coeff = (
|
450 |
+
alpha_prod_t_prev ** (0.5) * current_beta_t
|
451 |
+
) / beta_prod_t
|
452 |
+
current_sample_coeff = current_alpha_t ** (0.5) * beta_prod_t_prev / beta_prod_t
|
453 |
+
|
454 |
+
# 5. Compute predicted previous sample ยต_t
|
455 |
+
# See formula (7) from https://arxiv.org/pdf/2006.11239.pdf
|
456 |
+
pred_prev_sample = (
|
457 |
+
pred_original_sample_coeff * pred_original_sample
|
458 |
+
+ current_sample_coeff * sample
|
459 |
+
)
|
460 |
+
|
461 |
+
return pred_prev_sample
|
462 |
+
|
463 |
+
|
464 |
+
def normalize(
|
465 |
+
z_t,
|
466 |
+
i,
|
467 |
+
max_norm_zs,
|
468 |
+
):
|
469 |
+
max_norm = max_norm_zs[i]
|
470 |
+
if max_norm < 0:
|
471 |
+
return z_t, 1
|
472 |
+
|
473 |
+
norm = torch.norm(z_t)
|
474 |
+
if norm < max_norm:
|
475 |
+
return z_t, 1
|
476 |
+
|
477 |
+
coeff = max_norm / norm
|
478 |
+
z_t = z_t * coeff
|
479 |
+
return z_t, coeff
|
480 |
+
|
481 |
+
|
482 |
+
def find_index(timesteps, timestep):
|
483 |
+
for i, t in enumerate(timesteps):
|
484 |
+
if t == timestep:
|
485 |
+
return i
|
486 |
+
return -1
|
487 |
+
|
488 |
+
map_timpstep_to_index = {
|
489 |
+
torch.tensor(799): 0,
|
490 |
+
torch.tensor(599): 1,
|
491 |
+
torch.tensor(399): 2,
|
492 |
+
torch.tensor(199): 3,
|
493 |
+
torch.tensor(799, device='cuda:0'): 0,
|
494 |
+
torch.tensor(599, device='cuda:0'): 1,
|
495 |
+
torch.tensor(399, device='cuda:0'): 2,
|
496 |
+
torch.tensor(199, device='cuda:0'): 3,
|
497 |
+
}
|
498 |
+
|
499 |
+
def step_save_latents(
|
500 |
+
self,
|
501 |
+
model_output: torch.FloatTensor,
|
502 |
+
timestep: int,
|
503 |
+
sample: torch.FloatTensor,
|
504 |
+
eta: float = 0.0,
|
505 |
+
use_clipped_model_output: bool = False,
|
506 |
+
generator=None,
|
507 |
+
variance_noise: Optional[torch.FloatTensor] = None,
|
508 |
+
return_dict: bool = True,
|
509 |
+
):
|
510 |
+
# print(self._save_timesteps)
|
511 |
+
# timestep_index = map_timpstep_to_index[timestep]
|
512 |
+
# timestep_index = ((self._save_timesteps == timestep).nonzero(as_tuple=True)[0]).item()
|
513 |
+
timestep_index = self._save_timesteps.index(timestep) if not self.clean_step_run else -1
|
514 |
+
next_timestep_index = timestep_index + 1 if not self.clean_step_run else -1
|
515 |
+
u_hat_t = self.step_function(
|
516 |
+
model_output=model_output,
|
517 |
+
timestep=timestep,
|
518 |
+
sample=sample,
|
519 |
+
eta=eta,
|
520 |
+
use_clipped_model_output=use_clipped_model_output,
|
521 |
+
generator=generator,
|
522 |
+
variance_noise=variance_noise,
|
523 |
+
return_dict=False,
|
524 |
+
scheduler=self,
|
525 |
+
)
|
526 |
+
|
527 |
+
x_t_minus_1 = self.x_ts[next_timestep_index]
|
528 |
+
self.x_ts_c_hat.append(u_hat_t)
|
529 |
+
|
530 |
+
z_t = x_t_minus_1 - u_hat_t
|
531 |
+
self.latents.append(z_t)
|
532 |
+
|
533 |
+
z_t, _ = normalize(z_t, timestep_index, self._config.max_norm_zs)
|
534 |
+
|
535 |
+
x_t_minus_1_predicted = u_hat_t + z_t
|
536 |
+
|
537 |
+
if not return_dict:
|
538 |
+
return (x_t_minus_1_predicted,)
|
539 |
+
|
540 |
+
return DDIMSchedulerOutput(prev_sample=x_t_minus_1, pred_original_sample=None)
|
541 |
+
|
542 |
+
|
543 |
+
def step_use_latents(
|
544 |
+
self,
|
545 |
+
model_output: torch.FloatTensor,
|
546 |
+
timestep: int,
|
547 |
+
sample: torch.FloatTensor,
|
548 |
+
eta: float = 0.0,
|
549 |
+
use_clipped_model_output: bool = False,
|
550 |
+
generator=None,
|
551 |
+
variance_noise: Optional[torch.FloatTensor] = None,
|
552 |
+
return_dict: bool = True,
|
553 |
+
):
|
554 |
+
# timestep_index = ((self._save_timesteps == timestep).nonzero(as_tuple=True)[0]).item()
|
555 |
+
timestep_index = self._timesteps.index(timestep) if not self.clean_step_run else -1
|
556 |
+
next_timestep_index = (
|
557 |
+
timestep_index + 1 if not self.clean_step_run else -1
|
558 |
+
)
|
559 |
+
z_t = self.latents[next_timestep_index] # + 1 because latents[0] is X_T
|
560 |
+
|
561 |
+
_, normalize_coefficient = normalize(
|
562 |
+
z_t[0] if self._config.breakdown == "x_t_hat_c_with_zeros" else z_t,
|
563 |
+
timestep_index,
|
564 |
+
self._config.max_norm_zs,
|
565 |
+
)
|
566 |
+
|
567 |
+
if normalize_coefficient == 0:
|
568 |
+
eta = 0
|
569 |
+
|
570 |
+
# eta = normalize_coefficient
|
571 |
+
|
572 |
+
x_t_hat_c_hat = self.step_function(
|
573 |
+
model_output=model_output,
|
574 |
+
timestep=timestep,
|
575 |
+
sample=sample,
|
576 |
+
eta=eta,
|
577 |
+
use_clipped_model_output=use_clipped_model_output,
|
578 |
+
generator=generator,
|
579 |
+
variance_noise=variance_noise,
|
580 |
+
return_dict=False,
|
581 |
+
scheduler=self,
|
582 |
+
)
|
583 |
+
|
584 |
+
w1 = self._config.ws1[timestep_index]
|
585 |
+
w2 = self._config.ws2[timestep_index]
|
586 |
+
|
587 |
+
x_t_minus_1_exact = self.x_ts[next_timestep_index]
|
588 |
+
x_t_minus_1_exact = x_t_minus_1_exact.expand_as(x_t_hat_c_hat)
|
589 |
+
|
590 |
+
x_t_c_hat: torch.Tensor = self.x_ts_c_hat[next_timestep_index]
|
591 |
+
if self._config.breakdown == "x_t_c_hat":
|
592 |
+
raise NotImplementedError("breakdown x_t_c_hat not implemented yet")
|
593 |
+
|
594 |
+
# x_t_c_hat = x_t_c_hat.expand_as(x_t_hat_c_hat)
|
595 |
+
x_t_c = x_t_c_hat[0].expand_as(x_t_hat_c_hat)
|
596 |
+
|
597 |
+
# if self._config.breakdown == "x_t_c_hat":
|
598 |
+
# v1 = x_t_hat_c_hat - x_t_c_hat
|
599 |
+
# v2 = x_t_c_hat - x_t_c
|
600 |
+
if (
|
601 |
+
self._config.breakdown == "x_t_hat_c"
|
602 |
+
or self._config.breakdown == "x_t_hat_c_with_zeros"
|
603 |
+
):
|
604 |
+
zero_index_reconstruction = 1 if not self.time_measure_n else 0
|
605 |
+
edit_prompts_num = (
|
606 |
+
(model_output.size(0) - zero_index_reconstruction) // 3
|
607 |
+
if self._config.breakdown == "x_t_hat_c_with_zeros" and not self.p_to_p
|
608 |
+
else (model_output.size(0) - zero_index_reconstruction) // 2
|
609 |
+
)
|
610 |
+
x_t_hat_c_indices = (zero_index_reconstruction, edit_prompts_num + zero_index_reconstruction)
|
611 |
+
edit_images_indices = (
|
612 |
+
edit_prompts_num + zero_index_reconstruction,
|
613 |
+
(
|
614 |
+
model_output.size(0)
|
615 |
+
if self._config.breakdown == "x_t_hat_c"
|
616 |
+
else zero_index_reconstruction + 2 * edit_prompts_num
|
617 |
+
),
|
618 |
+
)
|
619 |
+
x_t_hat_c = torch.zeros_like(x_t_hat_c_hat)
|
620 |
+
x_t_hat_c[edit_images_indices[0] : edit_images_indices[1]] = x_t_hat_c_hat[
|
621 |
+
x_t_hat_c_indices[0] : x_t_hat_c_indices[1]
|
622 |
+
]
|
623 |
+
v1 = x_t_hat_c_hat - x_t_hat_c
|
624 |
+
v2 = x_t_hat_c - normalize_coefficient * x_t_c
|
625 |
+
if self._config.breakdown == "x_t_hat_c_with_zeros" and not self.p_to_p:
|
626 |
+
path = os.path.join(
|
627 |
+
self.folder_name,
|
628 |
+
VECTOR_DATA_FOLDER,
|
629 |
+
self.image_name,
|
630 |
+
)
|
631 |
+
if not hasattr(self, VECTOR_DATA_DICT):
|
632 |
+
os.makedirs(path, exist_ok=True)
|
633 |
+
self.vector_data = dict()
|
634 |
+
|
635 |
+
x_t_0 = x_t_c_hat[1]
|
636 |
+
empty_prompt_indices = (1 + 2 * edit_prompts_num, 1 + 3 * edit_prompts_num)
|
637 |
+
x_t_hat_0 = x_t_hat_c_hat[empty_prompt_indices[0] : empty_prompt_indices[1]]
|
638 |
+
|
639 |
+
self.vector_data[timestep.item()] = dict()
|
640 |
+
self.vector_data[timestep.item()]["x_t_hat_c"] = x_t_hat_c[
|
641 |
+
edit_images_indices[0] : edit_images_indices[1]
|
642 |
+
]
|
643 |
+
self.vector_data[timestep.item()]["x_t_hat_0"] = x_t_hat_0
|
644 |
+
self.vector_data[timestep.item()]["x_t_c"] = x_t_c[0].expand_as(x_t_hat_0)
|
645 |
+
self.vector_data[timestep.item()]["x_t_0"] = x_t_0.expand_as(x_t_hat_0)
|
646 |
+
self.vector_data[timestep.item()]["x_t_hat_c_hat"] = x_t_hat_c_hat[
|
647 |
+
edit_images_indices[0] : edit_images_indices[1]
|
648 |
+
]
|
649 |
+
self.vector_data[timestep.item()]["x_t_minus_1_noisy"] = x_t_minus_1_exact[
|
650 |
+
0
|
651 |
+
].expand_as(x_t_hat_0)
|
652 |
+
self.vector_data[timestep.item()]["x_t_minus_1_clean"] = self.x_0s[
|
653 |
+
next_timestep_index
|
654 |
+
].expand_as(x_t_hat_0)
|
655 |
+
|
656 |
+
else: # no breakdown
|
657 |
+
v1 = x_t_hat_c_hat - normalize_coefficient * x_t_c
|
658 |
+
v2 = 0
|
659 |
+
|
660 |
+
if self.save_intermediate_results and not self.p_to_p:
|
661 |
+
delta = v1 + v2
|
662 |
+
v1_plus_x0 = self.x_0s[next_timestep_index] + v1
|
663 |
+
v2_plus_x0 = self.x_0s[next_timestep_index] + v2
|
664 |
+
delta_plus_x0 = self.x_0s[next_timestep_index] + delta
|
665 |
+
|
666 |
+
v1_images = decode_latents(v1, self.pipe)
|
667 |
+
self.v1s_images.append(v1_images)
|
668 |
+
v2_images = (
|
669 |
+
decode_latents(v2, self.pipe)
|
670 |
+
if self._config.breakdown != "no_breakdown"
|
671 |
+
else [PIL.Image.new("RGB", (1, 1))]
|
672 |
+
)
|
673 |
+
self.v2s_images.append(v2_images)
|
674 |
+
delta_images = decode_latents(delta, self.pipe)
|
675 |
+
self.deltas_images.append(delta_images)
|
676 |
+
v1_plus_x0_images = decode_latents(v1_plus_x0, self.pipe)
|
677 |
+
self.v1_x0s.append(v1_plus_x0_images)
|
678 |
+
v2_plus_x0_images = (
|
679 |
+
decode_latents(v2_plus_x0, self.pipe)
|
680 |
+
if self._config.breakdown != "no_breakdown"
|
681 |
+
else [PIL.Image.new("RGB", (1, 1))]
|
682 |
+
)
|
683 |
+
self.v2_x0s.append(v2_plus_x0_images)
|
684 |
+
delta_plus_x0_images = decode_latents(delta_plus_x0, self.pipe)
|
685 |
+
self.deltas_x0s.append(delta_plus_x0_images)
|
686 |
+
|
687 |
+
# print(f"v1 norm: {torch.norm(v1, dim=0).mean()}")
|
688 |
+
# if self._config.breakdown != "no_breakdown":
|
689 |
+
# print(f"v2 norm: {torch.norm(v2, dim=0).mean()}")
|
690 |
+
# print(f"v sum norm: {torch.norm(v1 + v2, dim=0).mean()}")
|
691 |
+
|
692 |
+
x_t_minus_1 = normalize_coefficient * x_t_minus_1_exact + w1 * v1 + w2 * v2
|
693 |
+
|
694 |
+
if (
|
695 |
+
self._config.breakdown == "x_t_hat_c"
|
696 |
+
or self._config.breakdown == "x_t_hat_c_with_zeros"
|
697 |
+
):
|
698 |
+
x_t_minus_1[x_t_hat_c_indices[0] : x_t_hat_c_indices[1]] = x_t_minus_1[
|
699 |
+
edit_images_indices[0] : edit_images_indices[1]
|
700 |
+
] # update x_t_hat_c to be x_t_hat_c_hat
|
701 |
+
if self._config.breakdown == "x_t_hat_c_with_zeros" and not self.p_to_p:
|
702 |
+
x_t_minus_1[empty_prompt_indices[0] : empty_prompt_indices[1]] = (
|
703 |
+
x_t_minus_1[edit_images_indices[0] : edit_images_indices[1]]
|
704 |
+
)
|
705 |
+
self.vector_data[timestep.item()]["x_t_minus_1_edited"] = x_t_minus_1[
|
706 |
+
edit_images_indices[0] : edit_images_indices[1]
|
707 |
+
]
|
708 |
+
if timestep == self._timesteps[-1]:
|
709 |
+
torch.save(
|
710 |
+
self.vector_data,
|
711 |
+
os.path.join(
|
712 |
+
path,
|
713 |
+
f"{VECTOR_DATA_DICT}.pt",
|
714 |
+
),
|
715 |
+
)
|
716 |
+
# p_to_p_force_perfect_reconstruction
|
717 |
+
if not self.time_measure_n:
|
718 |
+
x_t_minus_1[0] = x_t_minus_1_exact[0]
|
719 |
+
|
720 |
+
if not return_dict:
|
721 |
+
return (x_t_minus_1,)
|
722 |
+
|
723 |
+
return DDIMSchedulerOutput(
|
724 |
+
prev_sample=x_t_minus_1,
|
725 |
+
pred_original_sample=None,
|
726 |
+
)
|
727 |
+
|
728 |
+
|
729 |
+
|
730 |
+
def get_ddpm_inversion_scheduler(
|
731 |
+
scheduler,
|
732 |
+
step_function,
|
733 |
+
config,
|
734 |
+
timesteps,
|
735 |
+
save_timesteps,
|
736 |
+
latents,
|
737 |
+
x_ts,
|
738 |
+
x_ts_c_hat,
|
739 |
+
save_intermediate_results,
|
740 |
+
pipe,
|
741 |
+
x_0,
|
742 |
+
v1s_images,
|
743 |
+
v2s_images,
|
744 |
+
deltas_images,
|
745 |
+
v1_x0s,
|
746 |
+
v2_x0s,
|
747 |
+
deltas_x0s,
|
748 |
+
folder_name,
|
749 |
+
image_name,
|
750 |
+
time_measure_n,
|
751 |
+
):
|
752 |
+
def step(
|
753 |
+
model_output: torch.FloatTensor,
|
754 |
+
timestep: int,
|
755 |
+
sample: torch.FloatTensor,
|
756 |
+
eta: float = 0.0,
|
757 |
+
use_clipped_model_output: bool = False,
|
758 |
+
generator=None,
|
759 |
+
variance_noise: Optional[torch.FloatTensor] = None,
|
760 |
+
return_dict: bool = True,
|
761 |
+
):
|
762 |
+
# if scheduler.is_save:
|
763 |
+
# start = timer()
|
764 |
+
res_inv = step_save_latents(
|
765 |
+
scheduler,
|
766 |
+
model_output[:1, :, :, :],
|
767 |
+
timestep,
|
768 |
+
sample[:1, :, :, :],
|
769 |
+
eta,
|
770 |
+
use_clipped_model_output,
|
771 |
+
generator,
|
772 |
+
variance_noise,
|
773 |
+
return_dict,
|
774 |
+
)
|
775 |
+
# end = timer()
|
776 |
+
# print(f"Run Time Inv: {end - start}")
|
777 |
+
|
778 |
+
res_inf = step_use_latents(
|
779 |
+
scheduler,
|
780 |
+
model_output[1:, :, :, :],
|
781 |
+
timestep,
|
782 |
+
sample[1:, :, :, :],
|
783 |
+
eta,
|
784 |
+
use_clipped_model_output,
|
785 |
+
generator,
|
786 |
+
variance_noise,
|
787 |
+
return_dict,
|
788 |
+
)
|
789 |
+
# res = res_inv
|
790 |
+
res = (torch.cat((res_inv[0], res_inf[0]), dim=0),)
|
791 |
+
return res
|
792 |
+
# return res
|
793 |
+
|
794 |
+
scheduler.step_function = step_function
|
795 |
+
scheduler.is_save = True
|
796 |
+
scheduler._timesteps = timesteps
|
797 |
+
scheduler._save_timesteps = save_timesteps if save_timesteps else timesteps
|
798 |
+
scheduler._config = config
|
799 |
+
scheduler.latents = latents
|
800 |
+
scheduler.x_ts = x_ts
|
801 |
+
scheduler.x_ts_c_hat = x_ts_c_hat
|
802 |
+
scheduler.step = step
|
803 |
+
scheduler.save_intermediate_results = save_intermediate_results
|
804 |
+
scheduler.pipe = pipe
|
805 |
+
scheduler.v1s_images = v1s_images
|
806 |
+
scheduler.v2s_images = v2s_images
|
807 |
+
scheduler.deltas_images = deltas_images
|
808 |
+
scheduler.v1_x0s = v1_x0s
|
809 |
+
scheduler.v2_x0s = v2_x0s
|
810 |
+
scheduler.deltas_x0s = deltas_x0s
|
811 |
+
scheduler.clean_step_run = False
|
812 |
+
scheduler.x_0s = create_xts(
|
813 |
+
config.noise_shift_delta,
|
814 |
+
config.noise_timesteps,
|
815 |
+
config.clean_step_timestep,
|
816 |
+
None,
|
817 |
+
pipe.scheduler,
|
818 |
+
timesteps,
|
819 |
+
x_0,
|
820 |
+
no_add_noise=True,
|
821 |
+
)
|
822 |
+
scheduler.folder_name = folder_name
|
823 |
+
scheduler.image_name = image_name
|
824 |
+
scheduler.p_to_p = False
|
825 |
+
scheduler.p_to_p_replace = False
|
826 |
+
scheduler.time_measure_n = time_measure_n
|
827 |
+
return scheduler
|
828 |
+
|
829 |
+
|
830 |
+
def create_grid(
|
831 |
+
images,
|
832 |
+
p_to_p_images,
|
833 |
+
prompts,
|
834 |
+
original_image_path,
|
835 |
+
):
|
836 |
+
images_len = len(images) if len(images) > 0 else len(p_to_p_images)
|
837 |
+
images_size = images[0].size if len(images) > 0 else p_to_p_images[0].size
|
838 |
+
x_0 = Image.open(original_image_path).resize(images_size)
|
839 |
+
|
840 |
+
images_ = [x_0] + images + ([x_0] + p_to_p_images if p_to_p_images else [])
|
841 |
+
|
842 |
+
l1 = 1 if len(images) > 0 else 0
|
843 |
+
l2 = 1 if len(p_to_p_images) else 0
|
844 |
+
grid = make_image_grid(images_, rows=l1 + l2, cols=images_len + 1, resize=None)
|
845 |
+
|
846 |
+
width = images_size[0]
|
847 |
+
height = width // 5
|
848 |
+
font = ImageFont.truetype("font.ttf", width // 14)
|
849 |
+
|
850 |
+
grid1 = Image.new("RGB", size=(grid.size[0], grid.size[1] + height))
|
851 |
+
grid1.paste(grid, (0, 0))
|
852 |
+
|
853 |
+
draw = ImageDraw.Draw(grid1)
|
854 |
+
|
855 |
+
c_width = 0
|
856 |
+
for prompt in prompts:
|
857 |
+
if len(prompt) > 30:
|
858 |
+
prompt = prompt[:30] + "\n" + prompt[30:]
|
859 |
+
draw.text((c_width, width * 2), prompt, font=font, fill=(255, 255, 255))
|
860 |
+
c_width += width
|
861 |
+
|
862 |
+
return grid1
|
863 |
+
|
864 |
+
|
865 |
+
def save_intermediate_results(
|
866 |
+
v1s_images,
|
867 |
+
v2s_images,
|
868 |
+
deltas_images,
|
869 |
+
v1_x0s,
|
870 |
+
v2_x0s,
|
871 |
+
deltas_x0s,
|
872 |
+
folder_name,
|
873 |
+
original_prompt,
|
874 |
+
):
|
875 |
+
from diffusers.utils import make_image_grid
|
876 |
+
|
877 |
+
path = f"{folder_name}/{original_prompt}_intermediate_results/"
|
878 |
+
os.makedirs(path, exist_ok=True)
|
879 |
+
make_image_grid(
|
880 |
+
list(itertools.chain(*v1s_images)),
|
881 |
+
rows=len(v1s_images),
|
882 |
+
cols=len(v1s_images[0]),
|
883 |
+
).save(f"{path}v1s_images.png")
|
884 |
+
make_image_grid(
|
885 |
+
list(itertools.chain(*v2s_images)),
|
886 |
+
rows=len(v2s_images),
|
887 |
+
cols=len(v2s_images[0]),
|
888 |
+
).save(f"{path}v2s_images.png")
|
889 |
+
make_image_grid(
|
890 |
+
list(itertools.chain(*deltas_images)),
|
891 |
+
rows=len(deltas_images),
|
892 |
+
cols=len(deltas_images[0]),
|
893 |
+
).save(f"{path}deltas_images.png")
|
894 |
+
make_image_grid(
|
895 |
+
list(itertools.chain(*v1_x0s)),
|
896 |
+
rows=len(v1_x0s),
|
897 |
+
cols=len(v1_x0s[0]),
|
898 |
+
).save(f"{path}v1_x0s.png")
|
899 |
+
make_image_grid(
|
900 |
+
list(itertools.chain(*v2_x0s)),
|
901 |
+
rows=len(v2_x0s),
|
902 |
+
cols=len(v2_x0s[0]),
|
903 |
+
).save(f"{path}v2_x0s.png")
|
904 |
+
make_image_grid(
|
905 |
+
list(itertools.chain(*deltas_x0s)),
|
906 |
+
rows=len(deltas_x0s[0]),
|
907 |
+
cols=len(deltas_x0s),
|
908 |
+
).save(f"{path}deltas_x0s.png")
|
909 |
+
for i, image in enumerate(list(itertools.chain(*deltas_x0s))):
|
910 |
+
image.save(f"{path}deltas_x0s_{i}.png")
|
911 |
+
|
912 |
+
|
913 |
+
# copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.py and removed the add_noise line
|
914 |
+
def prepare_latents_no_add_noise(
|
915 |
+
self,
|
916 |
+
image,
|
917 |
+
timestep,
|
918 |
+
batch_size,
|
919 |
+
num_images_per_prompt,
|
920 |
+
dtype,
|
921 |
+
device,
|
922 |
+
generator=None,
|
923 |
+
):
|
924 |
+
from diffusers.utils import deprecate
|
925 |
+
|
926 |
+
if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)):
|
927 |
+
raise ValueError(
|
928 |
+
f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
|
929 |
+
)
|
930 |
+
|
931 |
+
image = image.to(device=device, dtype=dtype)
|
932 |
+
|
933 |
+
batch_size = batch_size * num_images_per_prompt
|
934 |
+
|
935 |
+
if image.shape[1] == 4:
|
936 |
+
init_latents = image
|
937 |
+
|
938 |
+
else:
|
939 |
+
if isinstance(generator, list) and len(generator) != batch_size:
|
940 |
+
raise ValueError(
|
941 |
+
f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
|
942 |
+
f" size of {batch_size}. Make sure the batch size matches the length of the generators."
|
943 |
+
)
|
944 |
+
|
945 |
+
elif isinstance(generator, list):
|
946 |
+
init_latents = [
|
947 |
+
self.retrieve_latents(
|
948 |
+
self.vae.encode(image[i : i + 1]), generator=generator[i]
|
949 |
+
)
|
950 |
+
for i in range(batch_size)
|
951 |
+
]
|
952 |
+
init_latents = torch.cat(init_latents, dim=0)
|
953 |
+
else:
|
954 |
+
init_latents = self.retrieve_latents(
|
955 |
+
self.vae.encode(image), generator=generator
|
956 |
+
)
|
957 |
+
|
958 |
+
init_latents = self.vae.config.scaling_factor * init_latents
|
959 |
+
|
960 |
+
if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0:
|
961 |
+
# expand init_latents for batch_size
|
962 |
+
deprecation_message = (
|
963 |
+
f"You have passed {batch_size} text prompts (`prompt`), but only {init_latents.shape[0]} initial"
|
964 |
+
" images (`image`). Initial images are now duplicating to match the number of text prompts. Note"
|
965 |
+
" that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update"
|
966 |
+
" your script to pass as many initial images as text prompts to suppress this warning."
|
967 |
+
)
|
968 |
+
deprecate(
|
969 |
+
"len(prompt) != len(image)",
|
970 |
+
"1.0.0",
|
971 |
+
deprecation_message,
|
972 |
+
standard_warn=False,
|
973 |
+
)
|
974 |
+
additional_image_per_prompt = batch_size // init_latents.shape[0]
|
975 |
+
init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0)
|
976 |
+
elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0:
|
977 |
+
raise ValueError(
|
978 |
+
f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts."
|
979 |
+
)
|
980 |
+
else:
|
981 |
+
init_latents = torch.cat([init_latents], dim=0)
|
982 |
+
|
983 |
+
# get latents
|
984 |
+
latents = init_latents
|
985 |
+
|
986 |
+
return latents
|
987 |
+
|
988 |
+
|
989 |
+
# Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt
|
990 |
+
def encode_prompt_empty_prompt_zeros_sdxl(
|
991 |
+
self,
|
992 |
+
prompt: str,
|
993 |
+
prompt_2: Optional[str] = None,
|
994 |
+
device: Optional[torch.device] = None,
|
995 |
+
num_images_per_prompt: int = 1,
|
996 |
+
do_classifier_free_guidance: bool = True,
|
997 |
+
negative_prompt: Optional[str] = None,
|
998 |
+
negative_prompt_2: Optional[str] = None,
|
999 |
+
prompt_embeds: Optional[torch.FloatTensor] = None,
|
1000 |
+
negative_prompt_embeds: Optional[torch.FloatTensor] = None,
|
1001 |
+
pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
|
1002 |
+
negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
|
1003 |
+
lora_scale: Optional[float] = None,
|
1004 |
+
clip_skip: Optional[int] = None,
|
1005 |
+
):
|
1006 |
+
r"""
|
1007 |
+
Encodes the prompt into text encoder hidden states.
|
1008 |
+
|
1009 |
+
Args:
|
1010 |
+
prompt (`str` or `List[str]`, *optional*):
|
1011 |
+
prompt to be encoded
|
1012 |
+
prompt_2 (`str` or `List[str]`, *optional*):
|
1013 |
+
The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
|
1014 |
+
used in both text-encoders
|
1015 |
+
device: (`torch.device`):
|
1016 |
+
torch device
|
1017 |
+
num_images_per_prompt (`int`):
|
1018 |
+
number of images that should be generated per prompt
|
1019 |
+
do_classifier_free_guidance (`bool`):
|
1020 |
+
whether to use classifier free guidance or not
|
1021 |
+
negative_prompt (`str` or `List[str]`, *optional*):
|
1022 |
+
The prompt or prompts not to guide the image generation. If not defined, one has to pass
|
1023 |
+
`negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
|
1024 |
+
less than `1`).
|
1025 |
+
negative_prompt_2 (`str` or `List[str]`, *optional*):
|
1026 |
+
The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
|
1027 |
+
`text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
|
1028 |
+
prompt_embeds (`torch.FloatTensor`, *optional*):
|
1029 |
+
Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
|
1030 |
+
provided, text embeddings will be generated from `prompt` input argument.
|
1031 |
+
negative_prompt_embeds (`torch.FloatTensor`, *optional*):
|
1032 |
+
Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
|
1033 |
+
weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
|
1034 |
+
argument.
|
1035 |
+
pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
|
1036 |
+
Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
|
1037 |
+
If not provided, pooled text embeddings will be generated from `prompt` input argument.
|
1038 |
+
negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
|
1039 |
+
Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
|
1040 |
+
weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
|
1041 |
+
input argument.
|
1042 |
+
lora_scale (`float`, *optional*):
|
1043 |
+
A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
|
1044 |
+
clip_skip (`int`, *optional*):
|
1045 |
+
Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
|
1046 |
+
the output of the pre-final layer will be used for computing the prompt embeddings.
|
1047 |
+
"""
|
1048 |
+
device = device or self._execution_device
|
1049 |
+
|
1050 |
+
# set lora scale so that monkey patched LoRA
|
1051 |
+
# function of text encoder can correctly access it
|
1052 |
+
if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin):
|
1053 |
+
self._lora_scale = lora_scale
|
1054 |
+
|
1055 |
+
# dynamically adjust the LoRA scale
|
1056 |
+
if self.text_encoder is not None:
|
1057 |
+
if not USE_PEFT_BACKEND:
|
1058 |
+
adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
|
1059 |
+
else:
|
1060 |
+
scale_lora_layers(self.text_encoder, lora_scale)
|
1061 |
+
|
1062 |
+
if self.text_encoder_2 is not None:
|
1063 |
+
if not USE_PEFT_BACKEND:
|
1064 |
+
adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale)
|
1065 |
+
else:
|
1066 |
+
scale_lora_layers(self.text_encoder_2, lora_scale)
|
1067 |
+
|
1068 |
+
prompt = [prompt] if isinstance(prompt, str) else prompt
|
1069 |
+
|
1070 |
+
if prompt is not None:
|
1071 |
+
batch_size = len(prompt)
|
1072 |
+
else:
|
1073 |
+
batch_size = prompt_embeds.shape[0]
|
1074 |
+
|
1075 |
+
# Define tokenizers and text encoders
|
1076 |
+
tokenizers = (
|
1077 |
+
[self.tokenizer, self.tokenizer_2]
|
1078 |
+
if self.tokenizer is not None
|
1079 |
+
else [self.tokenizer_2]
|
1080 |
+
)
|
1081 |
+
text_encoders = (
|
1082 |
+
[self.text_encoder, self.text_encoder_2]
|
1083 |
+
if self.text_encoder is not None
|
1084 |
+
else [self.text_encoder_2]
|
1085 |
+
)
|
1086 |
+
|
1087 |
+
if prompt_embeds is None:
|
1088 |
+
prompt_2 = prompt_2 or prompt
|
1089 |
+
prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2
|
1090 |
+
|
1091 |
+
# textual inversion: procecss multi-vector tokens if necessary
|
1092 |
+
prompt_embeds_list = []
|
1093 |
+
prompts = [prompt, prompt_2]
|
1094 |
+
for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders):
|
1095 |
+
|
1096 |
+
text_inputs = tokenizer(
|
1097 |
+
prompt,
|
1098 |
+
padding="max_length",
|
1099 |
+
max_length=tokenizer.model_max_length,
|
1100 |
+
truncation=True,
|
1101 |
+
return_tensors="pt",
|
1102 |
+
)
|
1103 |
+
|
1104 |
+
text_input_ids = text_inputs.input_ids
|
1105 |
+
untruncated_ids = tokenizer(
|
1106 |
+
prompt, padding="longest", return_tensors="pt"
|
1107 |
+
).input_ids
|
1108 |
+
|
1109 |
+
if untruncated_ids.shape[-1] >= text_input_ids.shape[
|
1110 |
+
-1
|
1111 |
+
] and not torch.equal(text_input_ids, untruncated_ids):
|
1112 |
+
removed_text = tokenizer.batch_decode(
|
1113 |
+
untruncated_ids[:, tokenizer.model_max_length - 1 : -1]
|
1114 |
+
)
|
1115 |
+
logger.warning(
|
1116 |
+
"The following part of your input was truncated because CLIP can only handle sequences up to"
|
1117 |
+
f" {tokenizer.model_max_length} tokens: {removed_text}"
|
1118 |
+
)
|
1119 |
+
|
1120 |
+
prompt_embeds = text_encoder(
|
1121 |
+
text_input_ids.to(device), output_hidden_states=True
|
1122 |
+
)
|
1123 |
+
|
1124 |
+
# We are only ALWAYS interested in the pooled output of the final text encoder
|
1125 |
+
pooled_prompt_embeds = prompt_embeds[0]
|
1126 |
+
if clip_skip is None:
|
1127 |
+
prompt_embeds = prompt_embeds.hidden_states[-2]
|
1128 |
+
else:
|
1129 |
+
# "2" because SDXL always indexes from the penultimate layer.
|
1130 |
+
prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)]
|
1131 |
+
|
1132 |
+
if self.config.force_zeros_for_empty_prompt:
|
1133 |
+
prompt_embeds[[i for i in range(len(prompt)) if prompt[i] == ""]] = 0
|
1134 |
+
pooled_prompt_embeds[
|
1135 |
+
[i for i in range(len(prompt)) if prompt[i] == ""]
|
1136 |
+
] = 0
|
1137 |
+
|
1138 |
+
prompt_embeds_list.append(prompt_embeds)
|
1139 |
+
|
1140 |
+
prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
|
1141 |
+
|
1142 |
+
# get unconditional embeddings for classifier free guidance
|
1143 |
+
zero_out_negative_prompt = (
|
1144 |
+
negative_prompt is None and self.config.force_zeros_for_empty_prompt
|
1145 |
+
)
|
1146 |
+
if (
|
1147 |
+
do_classifier_free_guidance
|
1148 |
+
and negative_prompt_embeds is None
|
1149 |
+
and zero_out_negative_prompt
|
1150 |
+
):
|
1151 |
+
negative_prompt_embeds = torch.zeros_like(prompt_embeds)
|
1152 |
+
negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
|
1153 |
+
elif do_classifier_free_guidance and negative_prompt_embeds is None:
|
1154 |
+
negative_prompt = negative_prompt or ""
|
1155 |
+
negative_prompt_2 = negative_prompt_2 or negative_prompt
|
1156 |
+
|
1157 |
+
# normalize str to list
|
1158 |
+
negative_prompt = (
|
1159 |
+
batch_size * [negative_prompt]
|
1160 |
+
if isinstance(negative_prompt, str)
|
1161 |
+
else negative_prompt
|
1162 |
+
)
|
1163 |
+
negative_prompt_2 = (
|
1164 |
+
batch_size * [negative_prompt_2]
|
1165 |
+
if isinstance(negative_prompt_2, str)
|
1166 |
+
else negative_prompt_2
|
1167 |
+
)
|
1168 |
+
|
1169 |
+
uncond_tokens: List[str]
|
1170 |
+
if prompt is not None and type(prompt) is not type(negative_prompt):
|
1171 |
+
raise TypeError(
|
1172 |
+
f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
|
1173 |
+
f" {type(prompt)}."
|
1174 |
+
)
|
1175 |
+
elif batch_size != len(negative_prompt):
|
1176 |
+
raise ValueError(
|
1177 |
+
f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
|
1178 |
+
f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
|
1179 |
+
" the batch size of `prompt`."
|
1180 |
+
)
|
1181 |
+
else:
|
1182 |
+
uncond_tokens = [negative_prompt, negative_prompt_2]
|
1183 |
+
|
1184 |
+
negative_prompt_embeds_list = []
|
1185 |
+
for negative_prompt, tokenizer, text_encoder in zip(
|
1186 |
+
uncond_tokens, tokenizers, text_encoders
|
1187 |
+
):
|
1188 |
+
|
1189 |
+
max_length = prompt_embeds.shape[1]
|
1190 |
+
uncond_input = tokenizer(
|
1191 |
+
negative_prompt,
|
1192 |
+
padding="max_length",
|
1193 |
+
max_length=max_length,
|
1194 |
+
truncation=True,
|
1195 |
+
return_tensors="pt",
|
1196 |
+
)
|
1197 |
+
|
1198 |
+
negative_prompt_embeds = text_encoder(
|
1199 |
+
uncond_input.input_ids.to(device),
|
1200 |
+
output_hidden_states=True,
|
1201 |
+
)
|
1202 |
+
# We are only ALWAYS interested in the pooled output of the final text encoder
|
1203 |
+
negative_pooled_prompt_embeds = negative_prompt_embeds[0]
|
1204 |
+
negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
|
1205 |
+
|
1206 |
+
negative_prompt_embeds_list.append(negative_prompt_embeds)
|
1207 |
+
|
1208 |
+
negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)
|
1209 |
+
|
1210 |
+
if self.text_encoder_2 is not None:
|
1211 |
+
prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
|
1212 |
+
else:
|
1213 |
+
prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device)
|
1214 |
+
|
1215 |
+
bs_embed, seq_len, _ = prompt_embeds.shape
|
1216 |
+
# duplicate text embeddings for each generation per prompt, using mps friendly method
|
1217 |
+
prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
|
1218 |
+
prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
|
1219 |
+
|
1220 |
+
if do_classifier_free_guidance:
|
1221 |
+
# duplicate unconditional embeddings for each generation per prompt, using mps friendly method
|
1222 |
+
seq_len = negative_prompt_embeds.shape[1]
|
1223 |
+
|
1224 |
+
if self.text_encoder_2 is not None:
|
1225 |
+
negative_prompt_embeds = negative_prompt_embeds.to(
|
1226 |
+
dtype=self.text_encoder_2.dtype, device=device
|
1227 |
+
)
|
1228 |
+
else:
|
1229 |
+
negative_prompt_embeds = negative_prompt_embeds.to(
|
1230 |
+
dtype=self.unet.dtype, device=device
|
1231 |
+
)
|
1232 |
+
|
1233 |
+
negative_prompt_embeds = negative_prompt_embeds.repeat(
|
1234 |
+
1, num_images_per_prompt, 1
|
1235 |
+
)
|
1236 |
+
negative_prompt_embeds = negative_prompt_embeds.view(
|
1237 |
+
batch_size * num_images_per_prompt, seq_len, -1
|
1238 |
+
)
|
1239 |
+
|
1240 |
+
pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
|
1241 |
+
bs_embed * num_images_per_prompt, -1
|
1242 |
+
)
|
1243 |
+
if do_classifier_free_guidance:
|
1244 |
+
negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(
|
1245 |
+
1, num_images_per_prompt
|
1246 |
+
).view(bs_embed * num_images_per_prompt, -1)
|
1247 |
+
|
1248 |
+
if self.text_encoder is not None:
|
1249 |
+
if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
|
1250 |
+
# Retrieve the original scale by scaling back the LoRA layers
|
1251 |
+
unscale_lora_layers(self.text_encoder, lora_scale)
|
1252 |
+
|
1253 |
+
if self.text_encoder_2 is not None:
|
1254 |
+
if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
|
1255 |
+
# Retrieve the original scale by scaling back the LoRA layers
|
1256 |
+
unscale_lora_layers(self.text_encoder_2, lora_scale)
|
1257 |
+
|
1258 |
+
return (
|
1259 |
+
prompt_embeds,
|
1260 |
+
negative_prompt_embeds,
|
1261 |
+
pooled_prompt_embeds,
|
1262 |
+
negative_pooled_prompt_embeds,
|
1263 |
+
)
|
1264 |
+
|
1265 |
+
|
1266 |
+
def create_xts(
|
1267 |
+
noise_shift_delta,
|
1268 |
+
noise_timesteps,
|
1269 |
+
clean_step_timestep,
|
1270 |
+
generator,
|
1271 |
+
scheduler,
|
1272 |
+
timesteps,
|
1273 |
+
x_0,
|
1274 |
+
no_add_noise=False,
|
1275 |
+
):
|
1276 |
+
if noise_timesteps is None:
|
1277 |
+
noising_delta = noise_shift_delta * (timesteps[0] - timesteps[1])
|
1278 |
+
noise_timesteps = [timestep - int(noising_delta) for timestep in timesteps]
|
1279 |
+
|
1280 |
+
first_x_0_idx = len(noise_timesteps)
|
1281 |
+
for i in range(len(noise_timesteps)):
|
1282 |
+
if noise_timesteps[i] <= 0:
|
1283 |
+
first_x_0_idx = i
|
1284 |
+
break
|
1285 |
+
|
1286 |
+
noise_timesteps = noise_timesteps[:first_x_0_idx]
|
1287 |
+
|
1288 |
+
x_0_expanded = x_0.expand(len(noise_timesteps), -1, -1, -1)
|
1289 |
+
noise = (
|
1290 |
+
torch.randn(x_0_expanded.size(), generator=generator, device="cpu").to(
|
1291 |
+
x_0.device
|
1292 |
+
)
|
1293 |
+
if not no_add_noise
|
1294 |
+
else torch.zeros_like(x_0_expanded)
|
1295 |
+
)
|
1296 |
+
x_ts = scheduler.add_noise(
|
1297 |
+
x_0_expanded,
|
1298 |
+
noise,
|
1299 |
+
torch.IntTensor(noise_timesteps),
|
1300 |
+
)
|
1301 |
+
x_ts = [t.unsqueeze(dim=0) for t in list(x_ts)]
|
1302 |
+
x_ts += [x_0] * (len(timesteps) - first_x_0_idx)
|
1303 |
+
x_ts += [x_0]
|
1304 |
+
if clean_step_timestep > 0:
|
1305 |
+
x_ts += [x_0]
|
1306 |
+
return x_ts
|
1307 |
+
|
1308 |
+
|
1309 |
+
# Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise
|
1310 |
+
def add_noise(
|
1311 |
+
self,
|
1312 |
+
original_samples: torch.FloatTensor,
|
1313 |
+
noise: torch.FloatTensor,
|
1314 |
+
image_timesteps: torch.IntTensor,
|
1315 |
+
noise_timesteps: torch.IntTensor,
|
1316 |
+
) -> torch.FloatTensor:
|
1317 |
+
# Make sure alphas_cumprod and timestep have same device and dtype as original_samples
|
1318 |
+
# Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement
|
1319 |
+
# for the subsequent add_noise calls
|
1320 |
+
self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device)
|
1321 |
+
alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype)
|
1322 |
+
timesteps = timesteps.to(original_samples.device)
|
1323 |
+
|
1324 |
+
sqrt_alpha_prod = alphas_cumprod[image_timesteps] ** 0.5
|
1325 |
+
sqrt_alpha_prod = sqrt_alpha_prod.flatten()
|
1326 |
+
while len(sqrt_alpha_prod.shape) < len(original_samples.shape):
|
1327 |
+
sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)
|
1328 |
+
|
1329 |
+
sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[noise_timesteps]) ** 0.5
|
1330 |
+
sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
|
1331 |
+
while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape):
|
1332 |
+
sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)
|
1333 |
+
|
1334 |
+
noisy_samples = (
|
1335 |
+
sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise
|
1336 |
+
)
|
1337 |
+
return noisy_samples
|
1338 |
+
|
1339 |
+
|
1340 |
+
def make_image_grid(
|
1341 |
+
images: List[PIL.Image.Image], rows: int, cols: int, resize: int = None, size=None
|
1342 |
+
) -> PIL.Image.Image:
|
1343 |
+
"""
|
1344 |
+
Prepares a single grid of images. Useful for visualization purposes.
|
1345 |
+
"""
|
1346 |
+
assert len(images) == rows * cols
|
1347 |
+
|
1348 |
+
if resize is not None:
|
1349 |
+
images = [img.resize((resize, resize)) for img in images]
|
1350 |
+
|
1351 |
+
w, h = size
|
1352 |
+
grid = Image.new("RGB", size=(cols * w, rows * h))
|
1353 |
+
|
1354 |
+
for i, img in enumerate(images):
|
1355 |
+
grid.paste(img, box=(i % cols * w, i // cols * h))
|
1356 |
+
return grid
|