Create sdxl_base.yaml
Browse files- sdxl_base.yaml +86 -0
sdxl_base.yaml
ADDED
@@ -0,0 +1,86 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
model:
|
2 |
+
target: sgm.models.diffusion.DiffusionEngine
|
3 |
+
params:
|
4 |
+
scale_factor: 0.13025
|
5 |
+
disable_first_stage_autocast: True
|
6 |
+
|
7 |
+
network_config:
|
8 |
+
target: sgm.modules.diffusionmodules.openaimodel.UNetModel
|
9 |
+
params:
|
10 |
+
adm_in_channels: 2816
|
11 |
+
num_classes: sequential
|
12 |
+
use_checkpoint: True
|
13 |
+
in_channels: 4
|
14 |
+
out_channels: 4
|
15 |
+
model_channels: 320
|
16 |
+
attention_resolutions: [4, 2]
|
17 |
+
num_res_blocks: 2
|
18 |
+
channel_mult: [1, 2, 4]
|
19 |
+
num_head_channels: 64
|
20 |
+
use_spatial_transformer: True
|
21 |
+
use_linear_in_transformer: True
|
22 |
+
transformer_depth: [1, 2, 10] # note: the first is unused (due to attn_res starting at 2) 32, 16, 8 --> 64, 32, 16
|
23 |
+
context_dim: 2048
|
24 |
+
spatial_transformer_attn_type: softmax
|
25 |
+
legacy: False
|
26 |
+
|
27 |
+
conditioner_config:
|
28 |
+
target: sgm.modules.GeneralConditioner
|
29 |
+
params:
|
30 |
+
emb_models:
|
31 |
+
# crossattn cond
|
32 |
+
- is_trainable: False
|
33 |
+
input_key: prompts
|
34 |
+
target: sgm.encoders.FrozenCLIPEmbedder
|
35 |
+
params:
|
36 |
+
layer: hidden
|
37 |
+
layer_idx: 11
|
38 |
+
# crossattn and vector cond
|
39 |
+
- is_trainable: False
|
40 |
+
input_key: prompts
|
41 |
+
target: sgm.encoders.FrozenOpenCLIPEmbedder2
|
42 |
+
params:
|
43 |
+
arch: ViT-bigG-14
|
44 |
+
version: laion2b_s39b_b160k
|
45 |
+
freeze: True
|
46 |
+
layer: penultimate
|
47 |
+
always_return_pooled: True
|
48 |
+
legacy: False
|
49 |
+
# vector cond
|
50 |
+
- is_trainable: False
|
51 |
+
input_key: original_size_as_tuple
|
52 |
+
target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
|
53 |
+
params:
|
54 |
+
outdim: 256 # multiplied by two
|
55 |
+
# vector cond
|
56 |
+
- is_trainable: False
|
57 |
+
input_key: crop_coords_top_left
|
58 |
+
target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
|
59 |
+
params:
|
60 |
+
outdim: 256 # multiplied by two
|
61 |
+
# vector cond
|
62 |
+
- is_trainable: False
|
63 |
+
input_key: target_size_as_tuple
|
64 |
+
target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
|
65 |
+
params:
|
66 |
+
outdim: 256 # multiplied by two
|
67 |
+
|
68 |
+
first_stage_config:
|
69 |
+
target: sgm.models.autoencoder.AutoencoderKLInferenceWrapper
|
70 |
+
params:
|
71 |
+
embed_dim: 4
|
72 |
+
monitor: val/rec_loss
|
73 |
+
ddconfig:
|
74 |
+
attn_type: vanilla
|
75 |
+
double_z: true
|
76 |
+
z_channels: 4
|
77 |
+
resolution: 256
|
78 |
+
in_channels: 3
|
79 |
+
out_ch: 3
|
80 |
+
ch: 128
|
81 |
+
ch_mult: [1, 2, 4, 4]
|
82 |
+
num_res_blocks: 2
|
83 |
+
attn_resolutions: []
|
84 |
+
dropout: 0.0
|
85 |
+
lossconfig:
|
86 |
+
target: torch.nn.Identity
|