Fabrice-TIERCELIN commited on
Commit
2f37681
1 Parent(s): df25700

Upload SUPIR_v0.yaml

Browse files
Files changed (1) hide show
  1. SUPIR_v0.yaml +162 -0
SUPIR_v0.yaml ADDED
@@ -0,0 +1,162 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ model:
2
+ target: SUPIR.models.SUPIR_model.SUPIRModel
3
+ params:
4
+ ae_dtype: bf16
5
+ diffusion_dtype: fp16
6
+ scale_factor: 0.13025
7
+ disable_first_stage_autocast: True
8
+ network_wrapper: sgm.modules.diffusionmodules.wrappers.ControlWrapper
9
+
10
+ denoiser_config:
11
+ target: sgm.modules.diffusionmodules.denoiser.DiscreteDenoiserWithControl
12
+ params:
13
+ num_idx: 1000
14
+ weighting_config:
15
+ target: sgm.modules.diffusionmodules.denoiser_weighting.EpsWeighting
16
+ scaling_config:
17
+ target: sgm.modules.diffusionmodules.denoiser_scaling.EpsScaling
18
+ discretization_config:
19
+ target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization
20
+
21
+ control_stage_config:
22
+ target: SUPIR.modules.SUPIR_v0.GLVControl
23
+ params:
24
+ adm_in_channels: 2816
25
+ num_classes: sequential
26
+ use_checkpoint: True
27
+ in_channels: 4
28
+ out_channels: 4
29
+ model_channels: 320
30
+ attention_resolutions: [4, 2]
31
+ num_res_blocks: 2
32
+ channel_mult: [1, 2, 4]
33
+ num_head_channels: 64
34
+ use_spatial_transformer: True
35
+ use_linear_in_transformer: True
36
+ transformer_depth: [1, 2, 10] # note: the first is unused (due to attn_res starting at 2) 32, 16, 8 --> 64, 32, 16
37
+ # transformer_depth: [1, 1, 4]
38
+ context_dim: 2048
39
+ spatial_transformer_attn_type: softmax-xformers
40
+ legacy: False
41
+ input_upscale: 1
42
+
43
+ network_config:
44
+ target: SUPIR.modules.SUPIR_v0.LightGLVUNet
45
+ params:
46
+ mode: XL-base
47
+ project_type: ZeroSFT
48
+ project_channel_scale: 2
49
+ adm_in_channels: 2816
50
+ num_classes: sequential
51
+ use_checkpoint: True
52
+ in_channels: 4
53
+ out_channels: 4
54
+ model_channels: 320
55
+ attention_resolutions: [4, 2]
56
+ num_res_blocks: 2
57
+ channel_mult: [1, 2, 4]
58
+ num_head_channels: 64
59
+ use_spatial_transformer: True
60
+ use_linear_in_transformer: True
61
+ transformer_depth: [1, 2, 10] # note: the first is unused (due to attn_res starting at 2) 32, 16, 8 --> 64, 32, 16
62
+ context_dim: 2048
63
+ spatial_transformer_attn_type: softmax-xformers
64
+ legacy: False
65
+
66
+ conditioner_config:
67
+ target: sgm.modules.GeneralConditionerWithControl
68
+ params:
69
+ emb_models:
70
+ # crossattn cond
71
+ - is_trainable: False
72
+ input_key: txt
73
+ target: sgm.modules.encoders.modules.FrozenCLIPEmbedder
74
+ params:
75
+ layer: hidden
76
+ layer_idx: 11
77
+ # crossattn and vector cond
78
+ - is_trainable: False
79
+ input_key: txt
80
+ target: sgm.modules.encoders.modules.FrozenOpenCLIPEmbedder2
81
+ params:
82
+ arch: ViT-bigG-14
83
+ version: laion2b_s39b_b160k
84
+ freeze: True
85
+ layer: penultimate
86
+ always_return_pooled: True
87
+ legacy: False
88
+ # vector cond
89
+ - is_trainable: False
90
+ input_key: original_size_as_tuple
91
+ target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
92
+ params:
93
+ outdim: 256 # multiplied by two
94
+ # vector cond
95
+ - is_trainable: False
96
+ input_key: crop_coords_top_left
97
+ target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
98
+ params:
99
+ outdim: 256 # multiplied by two
100
+ # vector cond
101
+ - is_trainable: False
102
+ input_key: target_size_as_tuple
103
+ target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
104
+ params:
105
+ outdim: 256 # multiplied by two
106
+
107
+ first_stage_config:
108
+ target: sgm.models.autoencoder.AutoencoderKLInferenceWrapper
109
+ params:
110
+ ckpt_path: ~
111
+ embed_dim: 4
112
+ monitor: val/rec_loss
113
+ ddconfig:
114
+ attn_type: vanilla-xformers
115
+ double_z: true
116
+ z_channels: 4
117
+ resolution: 256
118
+ in_channels: 3
119
+ out_ch: 3
120
+ ch: 128
121
+ ch_mult: [ 1, 2, 4, 4 ]
122
+ num_res_blocks: 2
123
+ attn_resolutions: [ ]
124
+ dropout: 0.0
125
+ lossconfig:
126
+ target: torch.nn.Identity
127
+
128
+ sampler_config:
129
+ target: sgm.modules.diffusionmodules.sampling.RestoreEDMSampler
130
+ params:
131
+ num_steps: 100
132
+ restore_cfg: 4.0
133
+ s_churn: 0
134
+ s_noise: 1.003
135
+ discretization_config:
136
+ target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization
137
+ guider_config:
138
+ target: sgm.modules.diffusionmodules.guiders.LinearCFG
139
+ params:
140
+ scale: 7.5
141
+ scale_min: 4.0
142
+
143
+ p_p:
144
+ 'Cinematic, High Contrast, highly detailed, taken using a Canon EOS R camera,
145
+ hyper detailed photo - realistic maximum detail, 32k, Color Grading, ultra HD, extreme meticulous detailing,
146
+ skin pore detailing, hyper sharpness, perfect without deformations.'
147
+ n_p:
148
+ 'painting, oil painting, illustration, drawing, art, sketch, oil painting, cartoon, CG Style, 3D render,
149
+ unreal engine, blurring, dirty, messy, worst quality, low quality, frames, watermark, signature,
150
+ jpeg artifacts, deformed, lowres, over-smooth'
151
+
152
+ SDXL_CKPT: /opt/data/private/AIGC_pretrain/SDXL_cache/sd_xl_base_1.0_0.9vae.safetensors
153
+ SUPIR_CKPT_F: /opt/data/private/AIGC_pretrain/SUPIR_cache/SUPIR-v0F.ckpt
154
+ SUPIR_CKPT_Q: /opt/data/private/AIGC_pretrain/SUPIR_cache/SUPIR-v0Q.ckpt
155
+ SUPIR_CKPT: ~
156
+
157
+ default_setting:
158
+ s_cfg_Quality: 7.5
159
+ spt_linear_CFG_Quality: 4.0
160
+ s_cfg_Fidelity: 4.0
161
+ spt_linear_CFG_Fidelity: 1.0
162
+ edm_steps: 50