stable-diffusion-webui

1
model:
2
  target: sgm.models.diffusion.DiffusionEngine
3
  params:
4
    scale_factor: 0.13025
5
    disable_first_stage_autocast: True
6

7
    denoiser_config:
8
      target: sgm.modules.diffusionmodules.denoiser.DiscreteDenoiser
9
      params:
10
        num_idx: 1000
11

12
        weighting_config:
13
          target: sgm.modules.diffusionmodules.denoiser_weighting.EpsWeighting
14
        scaling_config:
15
          target: sgm.modules.diffusionmodules.denoiser_scaling.EpsScaling
16
        discretization_config:
17
          target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization
18

19
    network_config:
20
      target: sgm.modules.diffusionmodules.openaimodel.UNetModel
21
      params:
22
        adm_in_channels: 2816
23
        num_classes: sequential
24
        use_checkpoint: True
25
        in_channels: 9
26
        out_channels: 4
27
        model_channels: 320
28
        attention_resolutions: [4, 2]
29
        num_res_blocks: 2
30
        channel_mult: [1, 2, 4]
31
        num_head_channels: 64
32
        use_spatial_transformer: True
33
        use_linear_in_transformer: True
34
        transformer_depth: [1, 2, 10]  # note: the first is unused (due to attn_res starting at 2) 32, 16, 8 --> 64, 32, 16
35
        context_dim: 2048
36
        spatial_transformer_attn_type: softmax-xformers
37
        legacy: False
38

39
    conditioner_config:
40
      target: sgm.modules.GeneralConditioner
41
      params:
42
        emb_models:
43
          # crossattn cond
44
          - is_trainable: False
45
            input_key: txt
46
            target: sgm.modules.encoders.modules.FrozenCLIPEmbedder
47
            params:
48
              layer: hidden
49
              layer_idx: 11
50
          # crossattn and vector cond
51
          - is_trainable: False
52
            input_key: txt
53
            target: sgm.modules.encoders.modules.FrozenOpenCLIPEmbedder2
54
            params:
55
              arch: ViT-bigG-14
56
              version: laion2b_s39b_b160k
57
              freeze: True
58
              layer: penultimate
59
              always_return_pooled: True
60
              legacy: False
61
          # vector cond
62
          - is_trainable: False
63
            input_key: original_size_as_tuple
64
            target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
65
            params:
66
              outdim: 256  # multiplied by two
67
          # vector cond
68
          - is_trainable: False
69
            input_key: crop_coords_top_left
70
            target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
71
            params:
72
              outdim: 256  # multiplied by two
73
          # vector cond
74
          - is_trainable: False
75
            input_key: target_size_as_tuple
76
            target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
77
            params:
78
              outdim: 256  # multiplied by two
79

80
    first_stage_config:
81
      target: sgm.models.autoencoder.AutoencoderKLInferenceWrapper
82
      params:
83
        embed_dim: 4
84
        monitor: val/rec_loss
85
        ddconfig:
86
          attn_type: vanilla-xformers
87
          double_z: true
88
          z_channels: 4
89
          resolution: 256
90
          in_channels: 3
91
          out_ch: 3
92
          ch: 128
93
          ch_mult: [1, 2, 4, 4]
94
          num_res_blocks: 2
95
          attn_resolutions: []
96
          dropout: 0.0
97
        lossconfig:
98
          target: torch.nn.Identity
99