-
Notifications
You must be signed in to change notification settings - Fork 104
/
surf_ukiyo.yaml
88 lines (64 loc) · 2.89 KB
/
surf_ukiyo.yaml
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
pretrained_model_path: "./ckpt/stable-diffusion-v1-4"
dataset_config:
path: "data/style/surf"
prompt: "a man with round helmet surfing on a white wave in blue ocean with a rope"
n_sample_frame: 1
sampling_rate: 8
# use_invertion_latents: True
editing_config:
use_invertion_latents: true
use_inversion_attention: true
guidance_scale: 7.5
editing_prompts: [
a man with round helmet surfing on a white wave in blue ocean with a rope,
The Ukiyo-e style painting of a man with round helmet surfing on a white wave in blue ocean with a rope
]
p2p_config:
0:
# Whether to directly copy the cross attention from source
# True: directly copy, better for object replacement
# False: keep source attention, better for style
is_replace_controller: False
# Semantic preserving and replacement Debug me
cross_replace_steps:
default_: 0.8
# Source background structure preserving, in [0, 1].
# e.g., =0.6 Replace the first 60% steps self-attention
self_replace_steps: 0.8
# Amplify the target-words cross attention, larger value, more close to target
eq_params:
words: ["silver", "sculpture"]
values: [2,2]
# Target structure-divergence hyperparames
# If you change the shape of object better to use all three line, otherwise, no need.
# Without following three lines, all self-attention will be replaced
blend_words: [['cat',], ["cat",]]
blend_self_attention: True
# blend_latents: False # performance not so good in our case, need debug
blend_th: [2, 2]
# preserve source structure of blend_words , [0, 1]
# default is blend_th: [2, 2] # preserve all source self-attention
# blend_th : [0.0, 0.0], mask -> 1, use more att_replace, more generated attention, less source acttention
1:
is_replace_controller: False
cross_replace_steps:
default_: 0.9
self_replace_steps: 0.9
eq_params:
words: ["Ukiyo-e"]
values: [10, 10] # amplify attention to the word "tiger" by *2
clip_length: "${..dataset_config.n_sample_frame}"
sample_seeds: [0]
num_inference_steps: 50
prompt2prompt_edit: True
model_config:
# lora: 160
# temporal_downsample_time: 4
SparseCausalAttention_index: ['mid']
least_sc_channel: 640
# least_sc_channel: 100000
test_pipeline_config:
target: video_diffusion.pipelines.p2p_ddim_spatial_temporal.P2pDDIMSpatioTemporalPipeline
num_inference_steps: "${..validation_sample_logger.num_inference_steps}"
train_steps: 50
seed: 0