File size: 6,024 Bytes
b6af722
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

"""The default image and video tokenizer configs."""

from cosmos_predict1.tokenizer.modules import (
    ContinuousFormulation,
    Decoder3DType,
    DecoderType,
    DiscreteQuantizer,
    Encoder3DType,
    EncoderType,
)

continuous_image = dict(
    # The attention resolution for res blocks.
    attn_resolutions=[32],
    # The base number of channels.
    channels=128,
    # The channel multipler for each resolution.
    channels_mult=[2, 4, 4],
    dropout=0.0,
    in_channels=3,
    # The spatial compression ratio.
    spatial_compression=16,
    # The number of layers in each res block.
    num_res_blocks=2,
    out_channels=3,
    resolution=1024,
    patch_size=4,
    patch_method="haar",
    # The output latent dimension (channels).
    latent_channels=16,
    # The encoder output channels just before sampling.
    # Which is also the decoder's input channels.
    z_channels=16,
    # A factor over the z_channels, to get the total channels the encoder should output.
    # For a VAE for instance, we want to output the mean and variance, so we need 2 * z_channels.
    z_factor=1,
    name="CI",
    # What formulation to use, either "AE" or "VAE".
    # Chose VAE here, since the pre-trained ckpt were of a VAE formulation.
    formulation=ContinuousFormulation.AE.name,
    # Specify type of encoder ["Default", "LiteVAE"]
    encoder=EncoderType.Default.name,
    # Specify type of decoder ["Default"]
    decoder=DecoderType.Default.name,
)
continuous_image_8x8_360p = dict(continuous_image)
continuous_image_8x8_360p["patch_size"] = 2
continuous_image_8x8_360p["spatial_compression"] = 8

continuous_image_16x16_360p = dict(continuous_image)
continuous_image_16x16_360p["patch_size"] = 2
continuous_image_16x16_360p["spatial_compression"] = 16


discrete_image = dict(
    # The attention resolution for res blocks.
    attn_resolutions=[32],
    # The base number of channels.
    channels=128,
    # The channel multipler for each resolution.
    channels_mult=[2, 4, 4],
    dropout=0.0,
    in_channels=3,
    # The spatial compression ratio.
    spatial_compression=16,
    # The number of layers in each res block.
    num_res_blocks=2,
    out_channels=3,
    resolution=1024,
    patch_size=4,
    patch_method="haar",
    # The encoder output channels just before sampling.
    z_channels=256,
    # A factor over the z_channels, to get the total channels the encoder should output.
    # for discrete tokenization, often we directly use the vector, so z_factor=1.
    z_factor=1,
    # The quantizer of choice, VQ, LFQ, FSQ, or ResFSQ.
    quantizer=DiscreteQuantizer.FSQ.name,
    # The embedding dimension post-quantization, which is also the input channels of the decoder.
    # Which is also the output
    embedding_dim=6,
    # The number of levels to use for fine-scalar quantization.
    levels=[8, 8, 8, 5, 5, 5],
    # The number of quantizers to use for residual fine-scalar quantization.
    num_quantizers=4,
    name="DI",
    # Specify type of encoder ["Default", "LiteVAE"]
    encoder=EncoderType.Default.name,
    # Specify type of decoder ["Default"]
    decoder=DecoderType.Default.name,
)
discrete_image_8x8_360p = dict(discrete_image)
discrete_image_8x8_360p["patch_size"] = 2
discrete_image_8x8_360p["spatial_compression"] = 8

discrete_image_16x16_360p = dict(discrete_image)
discrete_image_16x16_360p["patch_size"] = 2
discrete_image_16x16_360p["spatial_compression"] = 16

continuous_video = dict(
    attn_resolutions=[32],
    channels=128,
    channels_mult=[2, 4, 4],
    dropout=0.0,
    in_channels=3,
    num_res_blocks=2,
    out_channels=3,
    resolution=1024,
    patch_size=4,
    patch_method="haar",
    latent_channels=16,
    z_channels=16,
    z_factor=1,
    num_groups=1,
    legacy_mode=False,
    spatial_compression=8,
    temporal_compression=8,
    formulation=ContinuousFormulation.AE.name,
    encoder=Encoder3DType.FACTORIZED.name,
    decoder=Decoder3DType.FACTORIZED.name,
    name="CV",
)

continuous_video_8x8x8_720p = dict(continuous_video)
continuous_video_8x8x8_720p["temporal_compression"] = 8
continuous_video_8x8x8_720p["spatial_compression"] = 8

continuous_video_4x8x8_360p = dict(continuous_video)
continuous_video_4x8x8_360p["temporal_compression"] = 4
continuous_video_4x8x8_360p["spatial_compression"] = 8
continuous_video_4x8x8_360p["patch_size"] = 2


discrete_video = dict(
    attn_resolutions=[32],
    channels=128,
    channels_mult=[2, 4, 4],
    dropout=0.0,
    in_channels=3,
    num_res_blocks=2,
    out_channels=3,
    resolution=1024,
    patch_size=4,
    patch_method="haar",
    z_channels=16,
    z_factor=1,
    num_groups=1,
    legacy_mode=False,
    spatial_compression=16,
    temporal_compression=8,
    quantizer=DiscreteQuantizer.FSQ.name,
    embedding_dim=6,
    levels=[8, 8, 8, 5, 5, 5],
    encoder=Encoder3DType.FACTORIZED.name,
    decoder=Decoder3DType.FACTORIZED.name,
    name="DV",
)

discrete_video_8x16x16_720p = dict(discrete_video)
discrete_video_8x16x16_720p["temporal_compression"] = 8
discrete_video_8x16x16_720p["spatial_compression"] = 16

discrete_video_4x8x8_360p = dict(discrete_video)
discrete_video_4x8x8_360p["z_channels"] = 256
discrete_video_4x8x8_360p["temporal_compression"] = 4
discrete_video_4x8x8_360p["spatial_compression"] = 8
discrete_video_4x8x8_360p["patch_size"] = 2