lev1 commited on
Commit
75453c0
1 Parent(s): 6a0dd6d

Demo version 3

Browse files
LICENSE ADDED
@@ -0,0 +1,83 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Copyright (c) 2023 Levon Khachatryan and Andranik Movsisyan and Vahram Tadevosyan and Roberto Henschel and Zhangyang Wang and Shant Navasardyan and Humphrey Shi
2
+
3
+
4
+ CreativeML Open RAIL-M
5
+ dated March 28, 2023
6
+
7
+ Section I: PREAMBLE
8
+
9
+ Multimodal generative models are being widely adopted and used, and have the potential to transform the way artists, among other individuals, conceive and benefit from AI or ML technologies as a tool for content creation.
10
+
11
+ Notwithstanding the current and potential benefits that these artifacts can bring to society at large, there are also concerns about potential misuses of them, either due to their technical limitations or ethical considerations.
12
+
13
+ In short, this license strives for both the open and responsible downstream use of the accompanying model. When it comes to the open character, we took inspiration from open source permissive licenses regarding the grant of IP rights. Referring to the downstream responsible use, we added use-based restrictions not permitting the use of the Model in very specific scenarios, in order for the licensor to be able to enforce the license in case potential misuses of the Model may occur. At the same time, we strive to promote open and responsible research on generative models for art and content generation.
14
+
15
+ Even though downstream derivative versions of the model could be released under different licensing terms, the latter will always have to include - at minimum - the same use-based restrictions as the ones in the original license (this license). We believe in the intersection between open and responsible AI development; thus, this License aims to strike a balance between both in order to enable responsible open-science in the field of AI.
16
+
17
+ This License governs the use of the model (and its derivatives) and is informed by the model card associated with the model.
18
+
19
+ NOW THEREFORE, You and Licensor agree as follows:
20
+
21
+ 1. Definitions
22
+
23
+ - "License" means the terms and conditions for use, reproduction, and Distribution as defined in this document.
24
+ - "Data" means a collection of information and/or content extracted from the dataset used with the Model, including to train, pretrain, or otherwise evaluate the Model. The Data is not licensed under this License.
25
+ - "Output" means the results of operating a Model as embodied in informational content resulting therefrom.
26
+ - "Model" means any accompanying machine-learning based assemblies (including checkpoints), consisting of learnt weights, parameters (including optimizer states), corresponding to the model architecture as embodied in the Complementary Material, that have been trained or tuned, in whole or in part on the Data, using the Complementary Material.
27
+ - "Derivatives of the Model" means all modifications to the Model, works based on the Model, or any other model which is created or initialized by transfer of patterns of the weights, parameters, activations or output of the Model, to the other model, in order to cause the other model to perform similarly to the Model, including - but not limited to - distillation methods entailing the use of intermediate data representations or methods based on the generation of synthetic data by the Model for training the other model.
28
+ - "Complementary Material" means the accompanying source code and scripts used to define, run, load, benchmark or evaluate the Model, and used to prepare data for training or evaluation, if any. This includes any accompanying documentation, tutorials, examples, etc, if any.
29
+ - "Distribution" means any transmission, reproduction, publication or other sharing of the Model or Derivatives of the Model to a third party, including providing the Model as a hosted service made available by electronic or other remote means - e.g. API-based or web access.
30
+ - "Licensor" means the copyright owner or entity authorized by the copyright owner that is granting the License, including the persons or entities that may have rights in the Model and/or distributing the Model.
31
+ - "You" (or "Your") means an individual or Legal Entity exercising permissions granted by this License and/or making use of the Model for whichever purpose and in any field of use, including usage of the Model in an end-use application - e.g. chatbot, translator, image generator.
32
+ - "Third Parties" means individuals or legal entities that are not under common control with Licensor or You.
33
+ - "Contribution" means any work of authorship, including the original version of the Model and any modifications or additions to that Model or Derivatives of the Model thereof, that is intentionally submitted to Licensor for inclusion in the Model by the copyright owner or by an individual or Legal Entity authorized to submit on behalf of the copyright owner. For the purposes of this definition, "submitted" means any form of electronic, verbal, or written communication sent to the Licensor or its representatives, including but not limited to communication on electronic mailing lists, source code control systems, and issue tracking systems that are managed by, or on behalf of, the Licensor for the purpose of discussing and improving the Model, but excluding communication that is conspicuously marked or otherwise designated in writing by the copyright owner as "Not a Contribution."
34
+ - "Contributor" means Licensor and any individual or Legal Entity on behalf of whom a Contribution has been received by Licensor and subsequently incorporated within the Model.
35
+
36
+ Section II: INTELLECTUAL PROPERTY RIGHTS
37
+
38
+ Both copyright and patent grants apply to the Model, Derivatives of the Model and Complementary Material. The Model and Derivatives of the Model are subject to additional terms as described in Section III.
39
+
40
+ 2. Grant of Copyright License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare, publicly display, publicly perform, sublicense, and distribute the Complementary Material, the Model, and Derivatives of the Model.
41
+ 3. Grant of Patent License. Subject to the terms and conditions of this License and where and as applicable, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable (except as stated in this paragraph) patent license to make, have made, use, offer to sell, sell, import, and otherwise transfer the Model and the Complementary Material, where such license applies only to those patent claims licensable by such Contributor that are necessarily infringed by their Contribution(s) alone or by combination of their Contribution(s) with the Model to which such Contribution(s) was submitted. If You institute patent litigation against any entity (including a cross-claim or counterclaim in a lawsuit) alleging that the Model and/or Complementary Material or a Contribution incorporated within the Model and/or Complementary Material constitutes direct or contributory patent infringement, then any patent licenses granted to You under this License for the Model and/or Work shall terminate as of the date such litigation is asserted or filed.
42
+
43
+ Section III: CONDITIONS OF USAGE, DISTRIBUTION AND REDISTRIBUTION
44
+
45
+ 4. Distribution and Redistribution. You may host for Third Party remote access purposes (e.g. software-as-a-service), reproduce and distribute copies of the Model or Derivatives of the Model thereof in any medium, with or without modifications, provided that You meet the following conditions:
46
+ Use-based restrictions as referenced in paragraph 5 MUST be included as an enforceable provision by You in any type of legal agreement (e.g. a license) governing the use and/or distribution of the Model or Derivatives of the Model, and You shall give notice to subsequent users You Distribute to, that the Model or Derivatives of the Model are subject to paragraph 5. This provision does not apply to the use of Complementary Material.
47
+ You must give any Third Party recipients of the Model or Derivatives of the Model a copy of this License;
48
+ You must cause any modified files to carry prominent notices stating that You changed the files;
49
+ You must retain all copyright, patent, trademark, and attribution notices excluding those notices that do not pertain to any part of the Model, Derivatives of the Model.
50
+ You may add Your own copyright statement to Your modifications and may provide additional or different license terms and conditions - respecting paragraph 4.a. - for use, reproduction, or Distribution of Your modifications, or for any such Derivatives of the Model as a whole, provided Your use, reproduction, and Distribution of the Model otherwise complies with the conditions stated in this License.
51
+ 5. Use-based restrictions. The restrictions set forth in Attachment A are considered Use-based restrictions. Therefore You cannot use the Model and the Derivatives of the Model for the specified restricted uses. You may use the Model subject to this License, including only for lawful purposes and in accordance with the License. Use may include creating any content with, finetuning, updating, running, training, evaluating and/or reparametrizing the Model. You shall require all of Your users who use the Model or a Derivative of the Model to comply with the terms of this paragraph (paragraph 5).
52
+ 6. The Output You Generate. Except as set forth herein, Licensor claims no rights in the Output You generate using the Model. You are accountable for the Output you generate and its subsequent uses. No use of the output can contravene any provision as stated in the License.
53
+
54
+ Section IV: OTHER PROVISIONS
55
+
56
+ 7. Updates and Runtime Restrictions. To the maximum extent permitted by law, Licensor reserves the right to restrict (remotely or otherwise) usage of the Model in violation of this License, update the Model through electronic means, or modify the Output of the Model based on updates. You shall undertake reasonable efforts to use the latest version of the Model.
57
+ 8. Trademarks and related. Nothing in this License permits You to make use of Licensors’ trademarks, trade names, logos or to otherwise suggest endorsement or misrepresent the relationship between the parties; and any rights not expressly granted herein are reserved by the Licensors.
58
+ 9. Disclaimer of Warranty. Unless required by applicable law or agreed to in writing, Licensor provides the Model and the Complementary Material (and each Contributor provides its Contributions) on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied, including, without limitation, any warranties or conditions of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A PARTICULAR PURPOSE. You are solely responsible for determining the appropriateness of using or redistributing the Model, Derivatives of the Model, and the Complementary Material and assume any risks associated with Your exercise of permissions under this License.
59
+ 10. Limitation of Liability. In no event and under no legal theory, whether in tort (including negligence), contract, or otherwise, unless required by applicable law (such as deliberate and grossly negligent acts) or agreed to in writing, shall any Contributor be liable to You for damages, including any direct, indirect, special, incidental, or consequential damages of any character arising as a result of this License or out of the use or inability to use the Model and the Complementary Material (including but not limited to damages for loss of goodwill, work stoppage, computer failure or malfunction, or any and all other commercial damages or losses), even if such Contributor has been advised of the possibility of such damages.
60
+ 11. Accepting Warranty or Additional Liability. While redistributing the Model, Derivatives of the Model and the Complementary Material thereof, You may choose to offer, and charge a fee for, acceptance of support, warranty, indemnity, or other liability obligations and/or rights consistent with this License. However, in accepting such obligations, You may act only on Your own behalf and on Your sole responsibility, not on behalf of any other Contributor, and only if You agree to indemnify, defend, and hold each Contributor harmless for any liability incurred by, or claims asserted against, such Contributor by reason of your accepting any such warranty or additional liability.
61
+ 12. If any provision of this License is held to be invalid, illegal or unenforceable, the remaining provisions shall be unaffected thereby and remain valid as if such provision had not been set forth herein.
62
+
63
+ END OF TERMS AND CONDITIONS
64
+
65
+
66
+
67
+
68
+ Attachment A
69
+
70
+ Use Restrictions
71
+
72
+ You agree not to use the Model or Derivatives of the Model:
73
+ - In any way that violates any applicable national, federal, state, local or international law or regulation;
74
+ - For the purpose of exploiting, harming or attempting to exploit or harm minors in any way;
75
+ - To generate or disseminate verifiably false information and/or content with the purpose of harming others;
76
+ - To generate or disseminate personal identifiable information that can be used to harm an individual;
77
+ - To defame, disparage or otherwise harass others;
78
+ - For fully automated decision making that adversely impacts an individual’s legal rights or otherwise creates or modifies a binding, enforceable obligation;
79
+ - For any use intended to or which has the effect of discriminating against or harming individuals or groups based on online or offline social behavior or known or predicted personal or personality characteristics;
80
+ - To exploit any of the vulnerabilities of a specific group of persons based on their age, social, physical or mental characteristics, in order to materially distort the behavior of a person pertaining to that group in a manner that causes or is likely to cause that person or another person physical or psychological harm;
81
+ - For any use intended to or which has the effect of discriminating against individuals or groups based on legally protected characteristics or categories;
82
+ - To provide medical advice and medical results interpretation;
83
+ - To generate or disseminate information for the purpose to be used for administration of justice, law enforcement, immigration or asylum processes, such as predicting an individual will commit fraud/crime commitment (e.g. by text profiling, drawing causal relationships between assertions made in documents, indiscriminate and arbitrarily-targeted use).
app_canny.py CHANGED
@@ -4,13 +4,13 @@ from model import Model
4
  def create_demo(model: Model):
5
 
6
  examples = [
7
- ["__assets__/canny_videos_edge/butterfly.mp4", "white butterfly, a high-quality, detailed, and professional photo"],
8
- ["__assets__/canny_videos_edge/deer.mp4", "oil painting of a deer, a high-quality, detailed, and professional photo"],
9
- ["__assets__/canny_videos_edge/fox.mp4", "wild red fox is walking on the grass, a high-quality, detailed, and professional photo"],
10
- ["__assets__/canny_videos_edge/girl_dancing.mp4", "oil painting of a girl dancing close-up, masterpiece, a high-quality, detailed, and professional photo"],
11
- ["__assets__/canny_videos_edge/girl_turning.mp4", "oil painting of a beautiful girl, a high-quality, detailed, and professional photo"],
12
- ["__assets__/canny_videos_edge/halloween.mp4", "beautiful girl halloween style, a high-quality, detailed, and professional photo"],
13
- ["__assets__/canny_videos_edge/santa.mp4", "a santa claus, a high-quality, detailed, and professional photo"],
14
  ]
15
 
16
  with gr.Blocks() as demo:
@@ -28,16 +28,22 @@ def create_demo(model: Model):
28
 
29
  with gr.Row():
30
  with gr.Column():
31
- input_video = gr.Video(label="Input Video",source='upload', format="mp4", visible=True).style(height="auto")
 
32
  with gr.Column():
33
  prompt = gr.Textbox(label='Prompt')
34
  run_button = gr.Button(label='Run')
 
 
 
35
  with gr.Column():
36
  result = gr.Video(label="Generated Video").style(height="auto")
37
 
38
  inputs = [
39
  input_video,
40
- prompt,
 
 
41
  ]
42
 
43
  gr.Examples(examples=examples,
 
4
  def create_demo(model: Model):
5
 
6
  examples = [
7
+ ["__assets__/canny_videos_edge_2fps/butterfly.mp4", "white butterfly, a high-quality, detailed, and professional photo"],
8
+ ["__assets__/canny_videos_edge_2fps/deer.mp4", "oil painting of a deer, a high-quality, detailed, and professional photo"],
9
+ ["__assets__/canny_videos_edge_2fps/fox.mp4", "wild red fox is walking on the grass, a high-quality, detailed, and professional photo"],
10
+ ["__assets__/canny_videos_edge_2fps/girl_dancing.mp4", "oil painting of a girl dancing close-up, masterpiece, a high-quality, detailed, and professional photo"],
11
+ ["__assets__/canny_videos_edge_2fps/girl_turning.mp4", "oil painting of a beautiful girl, a high-quality, detailed, and professional photo"],
12
+ ["__assets__/canny_videos_edge_2fps/halloween.mp4", "beautiful girl halloween style, a high-quality, detailed, and professional photo"],
13
+ ["__assets__/canny_videos_edge_2fps/santa.mp4", "a santa claus, a high-quality, detailed, and professional photo"],
14
  ]
15
 
16
  with gr.Blocks() as demo:
 
28
 
29
  with gr.Row():
30
  with gr.Column():
31
+ input_video = gr.Video(
32
+ label="Input Video", source='upload', format="mp4", visible=True).style(height="auto")
33
  with gr.Column():
34
  prompt = gr.Textbox(label='Prompt')
35
  run_button = gr.Button(label='Run')
36
+ with gr.Accordion('Advanced options', open=False):
37
+ watermark = gr.Radio(["Picsart AI Research", "Text2Video-Zero", "None"], label="Watermark", value='Picsart AI Research')
38
+ chunk_size = gr.Slider(label="Chunk size", minimum=2, maximum=8, value=8, step=1)
39
  with gr.Column():
40
  result = gr.Video(label="Generated Video").style(height="auto")
41
 
42
  inputs = [
43
  input_video,
44
+ prompt,
45
+ chunk_size,
46
+ watermark,
47
  ]
48
 
49
  gr.Examples(examples=examples,
app_canny_db.py CHANGED
@@ -38,11 +38,15 @@ def create_demo(model: Model):
38
 
39
  with gr.Row():
40
  with gr.Column():
 
41
  gr.Markdown("## Selection")
42
  db_text_field = gr.Markdown('DB Model: **Anime DB** ')
43
  canny_text_field = gr.Markdown('Motion: **woman1**')
44
  prompt = gr.Textbox(label='Prompt')
45
  run_button = gr.Button(label='Run')
 
 
 
46
  with gr.Column():
47
  result = gr.Image(label="Generated Video").style(height=400)
48
 
@@ -65,6 +69,8 @@ def create_demo(model: Model):
65
  db_selection,
66
  canny_selection,
67
  prompt,
 
 
68
  ]
69
 
70
  gr.Examples(examples=examples,
 
38
 
39
  with gr.Row():
40
  with gr.Column():
41
+ # input_video_path = gr.Video(source='upload', format="mp4", visible=False)
42
  gr.Markdown("## Selection")
43
  db_text_field = gr.Markdown('DB Model: **Anime DB** ')
44
  canny_text_field = gr.Markdown('Motion: **woman1**')
45
  prompt = gr.Textbox(label='Prompt')
46
  run_button = gr.Button(label='Run')
47
+ with gr.Accordion('Advanced options', open=False):
48
+ watermark = gr.Radio(["Picsart AI Research", "Text2Video-Zero", "None"], label="Watermark", value='Picsart AI Research')
49
+ chunk_size = gr.Slider(label="Chunk size", minimum=2, maximum=8, value=8, step=1)
50
  with gr.Column():
51
  result = gr.Image(label="Generated Video").style(height=400)
52
 
 
69
  db_selection,
70
  canny_selection,
71
  prompt,
72
+ chunk_size,
73
+ watermark,
74
  ]
75
 
76
  gr.Examples(examples=examples,
app_pix2pix_video.py CHANGED
@@ -4,10 +4,12 @@ from model import Model
4
 
5
  def create_demo(model: Model):
6
  examples = [
7
- ['__assets__/pix2pix video/camel.mp4', 'make it Van Gogh Starry Night style'],
8
- ['__assets__/pix2pix video/mini-cooper.mp4', 'make it Picasso style'],
9
- ['__assets__/pix2pix video/snowboard.mp4', 'replace man with robot'],
10
- ['__assets__/pix2pix video/white-swan.mp4', 'replace swan with mallard'],
 
 
11
  ]
12
  with gr.Blocks() as demo:
13
  with gr.Row():
@@ -29,6 +31,7 @@ def create_demo(model: Model):
29
  prompt = gr.Textbox(label='Prompt')
30
  run_button = gr.Button(label='Run')
31
  with gr.Accordion('Advanced options', open=False):
 
32
  image_resolution = gr.Slider(label='Image Resolution',
33
  minimum=256,
34
  maximum=1024,
@@ -39,31 +42,40 @@ def create_demo(model: Model):
39
  maximum=65536,
40
  value=0,
41
  step=1)
 
 
 
 
 
42
  start_t = gr.Slider(label='Starting time in seconds',
43
  minimum=0,
44
  maximum=10,
45
  value=0,
46
  step=1)
47
- end_t = gr.Slider(label='End time in seconds',
48
  minimum=0,
49
- maximum=15,
 
50
  step=1)
51
- out_fps = gr.Slider(label='Output video fps',
52
  minimum=1,
53
  maximum=30,
54
  value=-1,
55
  step=1)
 
56
  with gr.Column():
57
- result = gr.Video(label='Output',
58
- show_label=True)
59
  inputs = [
60
  input_image,
61
  prompt,
62
  image_resolution,
63
  seed,
 
64
  start_t,
65
  end_t,
66
- out_fps
 
 
67
  ]
68
 
69
  gr.Examples(examples=examples,
 
4
 
5
  def create_demo(model: Model):
6
  examples = [
7
+ ['__assets__/pix2pix_video_2fps/camel.mp4', 'make it Van Gogh Starry Night style', 512, 0, 1.0],
8
+ ['__assets__/pix2pix_video_2fps/mini-cooper.mp4', 'make it Picasso style', 512, 0, 1.5],
9
+ ['__assets__/pix2pix_video_2fps/snowboard.mp4', 'replace man with robot', 512, 0, 1.0],
10
+ ['__assets__/pix2pix_video_2fps/white-swan.mp4', 'replace swan with mallard', 512, 0, 1.5],
11
+ ['__assets__/pix2pix_video_2fps/boat.mp4', 'add city skyline in the background', 512, 0, 1.5],
12
+ ['__assets__/pix2pix_video_2fps/ballet.mp4', 'make her a golden sculpture', 512, 0, 1.0],
13
  ]
14
  with gr.Blocks() as demo:
15
  with gr.Row():
 
31
  prompt = gr.Textbox(label='Prompt')
32
  run_button = gr.Button(label='Run')
33
  with gr.Accordion('Advanced options', open=False):
34
+ watermark = gr.Radio(["Picsart AI Research", "Text2Video-Zero", "None"], label="Watermark", value='Picsart AI Research')
35
  image_resolution = gr.Slider(label='Image Resolution',
36
  minimum=256,
37
  maximum=1024,
 
42
  maximum=65536,
43
  value=0,
44
  step=1)
45
+ image_guidance = gr.Slider(label='Image guidance scale',
46
+ minimum=0.5,
47
+ maximum=2,
48
+ value=1.0,
49
+ step=0.1)
50
  start_t = gr.Slider(label='Starting time in seconds',
51
  minimum=0,
52
  maximum=10,
53
  value=0,
54
  step=1)
55
+ end_t = gr.Slider(label='End time in seconds (-1 corresponds to uploaded video duration)',
56
  minimum=0,
57
+ maximum=10,
58
+ value=-1,
59
  step=1)
60
+ out_fps = gr.Slider(label='Output video fps (-1 corresponds to uploaded video fps)',
61
  minimum=1,
62
  maximum=30,
63
  value=-1,
64
  step=1)
65
+ chunk_size = gr.Slider(label="Chunk size", minimum=2, maximum=8, value=8, step=1)
66
  with gr.Column():
67
+ result = gr.Video(label='Output', show_label=True)
 
68
  inputs = [
69
  input_image,
70
  prompt,
71
  image_resolution,
72
  seed,
73
+ image_guidance,
74
  start_t,
75
  end_t,
76
+ out_fps,
77
+ chunk_size,
78
+ watermark
79
  ]
80
 
81
  gr.Examples(examples=examples,
app_pose.py CHANGED
@@ -4,29 +4,20 @@ import os
4
  from model import Model
5
 
6
  examples = [
7
- ['Motion 1', "A Robot is dancing in Sahara desert"],
8
- ['Motion 2', "A Robot is dancing in Sahara desert"],
9
- ['Motion 3', "A Robot is dancing in Sahara desert"],
10
- ['Motion 4', "A Robot is dancing in Sahara desert"],
11
- ['Motion 5', "A Robot is dancing in Sahara desert"],
12
  ]
13
 
14
  def create_demo(model: Model):
15
  with gr.Blocks() as demo:
16
  with gr.Row():
17
  gr.Markdown('## Text and Pose Conditional Video Generation')
18
- # with gr.Row():
19
- # gr.HTML(
20
- # """
21
- # <div style="text-align: left; auto;">
22
- # <h2 style="font-weight: 450; font-size: 1rem; margin: 0rem">
23
- # Description:
24
- # </h3>
25
- # </div>
26
- # """)
27
 
28
  with gr.Row():
29
- gr.Markdown('### You must select one pose sequence shown on the right, or use the examples')
30
  with gr.Column():
31
  gallery_pose_sequence = gr.Gallery(label="Pose Sequence", value=[('__assets__/poses_skeleton_gifs/dance1.gif', "Motion 1"), ('__assets__/poses_skeleton_gifs/dance2.gif', "Motion 2"), ('__assets__/poses_skeleton_gifs/dance3.gif', "Motion 3"), ('__assets__/poses_skeleton_gifs/dance4.gif', "Motion 4"), ('__assets__/poses_skeleton_gifs/dance5.gif', "Motion 5")]).style(grid=[2], height="auto")
32
  input_video_path = gr.Textbox(label="Pose Sequence",visible=False,value="Motion 1")
@@ -35,6 +26,9 @@ def create_demo(model: Model):
35
  with gr.Column():
36
  prompt = gr.Textbox(label='Prompt')
37
  run_button = gr.Button(label='Run')
 
 
 
38
  with gr.Column():
39
  result = gr.Image(label="Generated Video")
40
 
@@ -43,6 +37,8 @@ def create_demo(model: Model):
43
  inputs = [
44
  input_video_path,
45
  prompt,
 
 
46
  ]
47
 
48
  gr.Examples(examples=examples,
@@ -61,7 +57,8 @@ def create_demo(model: Model):
61
 
62
 
63
  def on_video_path_update(evt: gr.EventData):
64
- return f'Pose Sequence: **{evt._data}**'
 
65
 
66
  def pose_gallery_callback(evt: gr.SelectData):
67
- return f"Motion {evt.index+1}"
 
4
  from model import Model
5
 
6
  examples = [
7
+ ['Motion 1', "An astronaut dancing in the outer space"],
8
+ ['Motion 2', "An astronaut dancing in the outer space"],
9
+ ['Motion 3', "An astronaut dancing in the outer space"],
10
+ ['Motion 4', "An astronaut dancing in the outer space"],
11
+ ['Motion 5', "An astronaut dancing in the outer space"],
12
  ]
13
 
14
  def create_demo(model: Model):
15
  with gr.Blocks() as demo:
16
  with gr.Row():
17
  gr.Markdown('## Text and Pose Conditional Video Generation')
 
 
 
 
 
 
 
 
 
18
 
19
  with gr.Row():
20
+ gr.Markdown('Selection: **one motion** and a **prompt**, or use the examples below.')
21
  with gr.Column():
22
  gallery_pose_sequence = gr.Gallery(label="Pose Sequence", value=[('__assets__/poses_skeleton_gifs/dance1.gif', "Motion 1"), ('__assets__/poses_skeleton_gifs/dance2.gif', "Motion 2"), ('__assets__/poses_skeleton_gifs/dance3.gif', "Motion 3"), ('__assets__/poses_skeleton_gifs/dance4.gif', "Motion 4"), ('__assets__/poses_skeleton_gifs/dance5.gif', "Motion 5")]).style(grid=[2], height="auto")
23
  input_video_path = gr.Textbox(label="Pose Sequence",visible=False,value="Motion 1")
 
26
  with gr.Column():
27
  prompt = gr.Textbox(label='Prompt')
28
  run_button = gr.Button(label='Run')
29
+ with gr.Accordion('Advanced options', open=False):
30
+ watermark = gr.Radio(["Picsart AI Research", "Text2Video-Zero", "None"], label="Watermark",value='Picsart AI Research')
31
+ chunk_size = gr.Slider(label="Chunk size", minimum=2, maximum=8, value=8, step=1)
32
  with gr.Column():
33
  result = gr.Image(label="Generated Video")
34
 
 
37
  inputs = [
38
  input_video_path,
39
  prompt,
40
+ chunk_size,
41
+ watermark,
42
  ]
43
 
44
  gr.Examples(examples=examples,
 
57
 
58
 
59
  def on_video_path_update(evt: gr.EventData):
60
+ return f'Selection: **{evt._data}**'
61
+
62
 
63
  def pose_gallery_callback(evt: gr.SelectData):
64
+ return f"Motion {evt.index+1}"
app_text_to_video.py CHANGED
@@ -1,6 +1,8 @@
1
  import gradio as gr
2
  from model import Model
3
  from functools import partial
 
 
4
 
5
  examples = [
6
  ["an astronaut waving the arm on the moon"],
@@ -15,6 +17,38 @@ examples = [
15
  ]
16
 
17
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
18
  def create_demo(model: Model):
19
 
20
  with gr.Blocks() as demo:
@@ -25,37 +59,47 @@ def create_demo(model: Model):
25
  """
26
  <div style="text-align: left; auto;">
27
  <h2 style="font-weight: 450; font-size: 1rem; margin: 0rem">
28
- Description: Simply input <b>any textual prompt</b> to generate videos right away and unleash your creativity and imagination! You can also select from the examples below. For performance purposes, our current preview release generates only 8 output frames and output 4s videos.
29
  </h3>
30
  </div>
31
  """)
32
 
33
  with gr.Row():
34
  with gr.Column():
 
 
 
 
 
35
  prompt = gr.Textbox(label='Prompt')
36
  run_button = gr.Button(label='Run')
37
  with gr.Accordion('Advanced options', open=False):
38
- motion_field_strength_x = gr.Slider(label='Global Translation $\delta_{x}$',
39
- minimum=-20,
40
- maximum=20,
41
- value=12,
42
- step=1)
43
-
44
- motion_field_strength_y = gr.Slider(label='Global Translation $\delta_{y}$',
45
- minimum=-20,
46
- maximum=20,
47
- value=12,
48
- step=1)
49
- # a_prompt = gr.Textbox(label="Added Prompt", value='best quality, extremely detailed')
50
- n_prompt = gr.Textbox(label="Optional Negative Prompt",
51
- value='')
52
  with gr.Column():
53
  result = gr.Video(label="Generated Video")
 
54
  inputs = [
55
  prompt,
 
56
  motion_field_strength_x,
57
  motion_field_strength_y,
58
- n_prompt
 
 
 
 
 
59
  ]
60
 
61
  gr.Examples(examples=examples,
 
1
  import gradio as gr
2
  from model import Model
3
  from functools import partial
4
+ from bs4 import BeautifulSoup
5
+ import requests
6
 
7
  examples = [
8
  ["an astronaut waving the arm on the moon"],
 
17
  ]
18
 
19
 
20
+ def model_url_list():
21
+ url_list = []
22
+ for i in range(0, 5):
23
+ url_list.append(f"https://huggingface.co/models?p={i}&sort=downloads&search=dreambooth")
24
+ return url_list
25
+
26
+ def data_scraping(url_list):
27
+ model_list = []
28
+ for url in url_list:
29
+ response = requests.get(url)
30
+ soup = BeautifulSoup(response.text, "html.parser")
31
+ div_class = 'grid grid-cols-1 gap-5 2xl:grid-cols-2'
32
+ div = soup.find('div', {'class': div_class})
33
+ for a in div.find_all('a', href=True):
34
+ model_list.append(a['href'])
35
+ return model_list
36
+
37
+ model_list = data_scraping(model_url_list())
38
+ for i in range(len(model_list)):
39
+ model_list[i] = model_list[i][1:]
40
+
41
+ best_model_list = [
42
+ "dreamlike-art/dreamlike-photoreal-2.0",
43
+ "dreamlike-art/dreamlike-diffusion-1.0",
44
+ "runwayml/stable-diffusion-v1-5",
45
+ "CompVis/stable-diffusion-v1-4",
46
+ "prompthero/openjourney",
47
+ ]
48
+
49
+ model_list = best_model_list + model_list
50
+
51
+
52
  def create_demo(model: Model):
53
 
54
  with gr.Blocks() as demo:
 
59
  """
60
  <div style="text-align: left; auto;">
61
  <h2 style="font-weight: 450; font-size: 1rem; margin: 0rem">
62
+ Description: Simply input <b>any textual prompt</b> to generate videos right away and unleash your creativity and imagination! You can also select from the examples below. For performance purposes, our current preview release by default generates only 8 output frames and output 4s videos, but you can increase it from Advanced Options.
63
  </h3>
64
  </div>
65
  """)
66
 
67
  with gr.Row():
68
  with gr.Column():
69
+ model_name = gr.Dropdown(
70
+ label="Model",
71
+ choices=model_list,
72
+ value="dreamlike-art/dreamlike-photoreal-2.0",
73
+ )
74
  prompt = gr.Textbox(label='Prompt')
75
  run_button = gr.Button(label='Run')
76
  with gr.Accordion('Advanced options', open=False):
77
+ watermark = gr.Radio(["Picsart AI Research", "Text2Video-Zero", "None"], label="Watermark", value='Picsart AI Research')
78
+
79
+ video_length = gr.Number(label="Video length", value=8, min=2, precision=0)
80
+ chunk_size = gr.Slider(label="Chunk size", minimum=2, maximum=32, value=8, step=1)
81
+
82
+ motion_field_strength_x = gr.Slider(label='Global Translation $\delta_{x}$', minimum=-20, maximum=20, value=12, step=1)
83
+ motion_field_strength_y = gr.Slider(label='Global Translation $\delta_{y}$', minimum=-20, maximum=20, value=12, step=1)
84
+
85
+ t0 = gr.Slider(label="Timestep t0", minimum=0, maximum=49, value=44, step=1)
86
+ t1 = gr.Slider(label="Timestep t1", minimum=0, maximum=49, value=47, step=1)
87
+
88
+ n_prompt = gr.Textbox(label="Optional Negative Prompt", value='')
 
 
89
  with gr.Column():
90
  result = gr.Video(label="Generated Video")
91
+
92
  inputs = [
93
  prompt,
94
+ model_name,
95
  motion_field_strength_x,
96
  motion_field_strength_y,
97
+ t0,
98
+ t1,
99
+ n_prompt,
100
+ chunk_size,
101
+ video_length,
102
+ watermark,
103
  ]
104
 
105
  gr.Examples(examples=examples,
gradio_utils.py CHANGED
@@ -4,23 +4,22 @@ def edge_path_to_video_path(edge_path):
4
 
5
  vid_name = edge_path.split("/")[-1]
6
  if vid_name == "butterfly.mp4":
7
- video_path = "__assets__/canny_videos_mp4/butterfly.mp4"
8
  elif vid_name == "deer.mp4":
9
- video_path = "__assets__/canny_videos_mp4/deer.mp4"
10
  elif vid_name == "fox.mp4":
11
- video_path = "__assets__/canny_videos_mp4/fox.mp4"
12
  elif vid_name == "girl_dancing.mp4":
13
- video_path = "__assets__/canny_videos_mp4/girl_dancing.mp4"
14
  elif vid_name == "girl_turning.mp4":
15
- video_path = "__assets__/canny_videos_mp4/girl_turning.mp4"
16
  elif vid_name == "halloween.mp4":
17
- video_path = "__assets__/canny_videos_mp4/halloween.mp4"
18
  elif vid_name == "santa.mp4":
19
- video_path = "__assets__/canny_videos_mp4/santa.mp4"
20
  return video_path
21
 
22
 
23
- # App Pose utils
24
  def motion_to_video_path(motion):
25
  videos = [
26
  "__assets__/poses_skeleton_gifs/dance1_corr.mp4",
@@ -29,23 +28,26 @@ def motion_to_video_path(motion):
29
  "__assets__/poses_skeleton_gifs/dance4_corr.mp4",
30
  "__assets__/poses_skeleton_gifs/dance5_corr.mp4"
31
  ]
32
- id = int(motion.split(" ")[1]) - 1
33
- return videos[id]
 
 
 
34
 
35
 
36
  # App Canny Dreambooth utils
37
  def get_video_from_canny_selection(canny_selection):
38
  if canny_selection == "woman1":
39
- input_video_path = "__assets__/db_files/woman1.mp4"
40
 
41
  elif canny_selection == "woman2":
42
- input_video_path = "__assets__/db_files/woman2.mp4"
43
 
44
  elif canny_selection == "man1":
45
- input_video_path = "__assets__/db_files/man1.mp4"
46
 
47
  elif canny_selection == "woman3":
48
- input_video_path = "__assets__/db_files/woman3.mp4"
49
  else:
50
  raise Exception
51
 
@@ -75,3 +77,13 @@ def get_canny_name_from_id(id):
75
  canny_names = ["woman1", "woman2", "man1", "woman3"]
76
  return canny_names[id]
77
 
 
 
 
 
 
 
 
 
 
 
 
4
 
5
  vid_name = edge_path.split("/")[-1]
6
  if vid_name == "butterfly.mp4":
7
+ video_path = "__assets__/canny_videos_mp4_2fps/butterfly.mp4"
8
  elif vid_name == "deer.mp4":
9
+ video_path = "__assets__/canny_videos_mp4_2fps/deer.mp4"
10
  elif vid_name == "fox.mp4":
11
+ video_path = "__assets__/canny_videos_mp4_2fps/fox.mp4"
12
  elif vid_name == "girl_dancing.mp4":
13
+ video_path = "__assets__/canny_videos_mp4_2fps/girl_dancing.mp4"
14
  elif vid_name == "girl_turning.mp4":
15
+ video_path = "__assets__/canny_videos_mp4_2fps/girl_turning.mp4"
16
  elif vid_name == "halloween.mp4":
17
+ video_path = "__assets__/canny_videos_mp4_2fps/halloween.mp4"
18
  elif vid_name == "santa.mp4":
19
+ video_path = "__assets__/canny_videos_mp4_2fps/santa.mp4"
20
  return video_path
21
 
22
 
 
23
  def motion_to_video_path(motion):
24
  videos = [
25
  "__assets__/poses_skeleton_gifs/dance1_corr.mp4",
 
28
  "__assets__/poses_skeleton_gifs/dance4_corr.mp4",
29
  "__assets__/poses_skeleton_gifs/dance5_corr.mp4"
30
  ]
31
+ if len(motion.split(" ")) > 1 and motion.split(" ")[1].isnumeric():
32
+ id = int(motion.split(" ")[1]) - 1
33
+ return videos[id]
34
+ else:
35
+ return motion
36
 
37
 
38
  # App Canny Dreambooth utils
39
  def get_video_from_canny_selection(canny_selection):
40
  if canny_selection == "woman1":
41
+ input_video_path = "__assets__/db_files_2fps/woman1.mp4"
42
 
43
  elif canny_selection == "woman2":
44
+ input_video_path = "__assets__/db_files_2fps/woman2.mp4"
45
 
46
  elif canny_selection == "man1":
47
+ input_video_path = "__assets__/db_files_2fps/man1.mp4"
48
 
49
  elif canny_selection == "woman3":
50
+ input_video_path = "__assets__/db_files_2fps/woman3.mp4"
51
  else:
52
  raise Exception
53
 
 
77
  canny_names = ["woman1", "woman2", "man1", "woman3"]
78
  return canny_names[id]
79
 
80
+
81
+ def logo_name_to_path(name):
82
+ logo_paths = {
83
+ 'Picsart AI Research': '__assets__/pair_watermark.png',
84
+ 'Text2Video-Zero': '__assets__/t2v-z_watermark.png',
85
+ 'None': None
86
+ }
87
+ if name in logo_paths:
88
+ return logo_paths[name]
89
+ return name
model.py CHANGED
@@ -11,7 +11,7 @@ from text_to_video.text_to_video_pipeline import TextToVideoPipeline
11
  import utils
12
  import gradio_utils
13
 
14
- decord.bridge.set_bridge('torch')
15
 
16
 
17
  class ModelType(Enum):
@@ -55,14 +55,19 @@ class Model:
55
  def inference_chunk(self, frame_ids, **kwargs):
56
  if self.pipe is None:
57
  return
58
- image = kwargs.pop('image')
59
  prompt = np.array(kwargs.pop('prompt'))
60
  negative_prompt = np.array(kwargs.pop('negative_prompt', ''))
61
  latents = None
62
  if 'latents' in kwargs:
63
  latents = kwargs.pop('latents')[frame_ids]
64
- return self.pipe(image=image[frame_ids],
65
- prompt=prompt[frame_ids].tolist(),
 
 
 
 
 
66
  negative_prompt=negative_prompt[frame_ids].tolist(),
67
  latents=latents,
68
  generator=self.generator,
@@ -72,15 +77,21 @@ class Model:
72
  if self.pipe is None:
73
  return
74
  seed = kwargs.pop('seed', 0)
 
 
75
  kwargs.pop('generator', '')
76
- # self.generator.manual_seed(seed)
 
 
 
 
 
 
 
 
 
 
77
  if split_to_chunks:
78
- assert 'image' in kwargs
79
- assert 'prompt' in kwargs
80
- image = kwargs.pop('image')
81
- prompt = kwargs.pop('prompt')
82
- negative_prompt = kwargs.pop('negative_prompt', '')
83
- f = image.shape[0]
84
  chunk_ids = np.arange(0, f, chunk_size - 1)
85
  result = []
86
  for i in range(len(chunk_ids)):
@@ -90,18 +101,19 @@ class Model:
90
  self.generator.manual_seed(seed)
91
  print(f'Processing chunk {i + 1} / {len(chunk_ids)}')
92
  result.append(self.inference_chunk(frame_ids=frame_ids,
93
- image=image,
94
- prompt=[prompt] * f,
95
- negative_prompt=[negative_prompt] * f,
96
  **kwargs).images[1:])
97
  result = np.concatenate(result)
98
  return result
99
  else:
100
- return self.pipe(generator=self.generator, **kwargs).videos[0]
101
 
102
  def process_controlnet_canny(self,
103
  video_path,
104
  prompt,
 
 
105
  num_inference_steps=20,
106
  controlnet_conditioning_scale=1.0,
107
  guidance_scale=9.0,
@@ -109,24 +121,32 @@ class Model:
109
  eta=0.0,
110
  low_threshold=100,
111
  high_threshold=200,
112
- resolution=512):
 
 
113
  video_path = gradio_utils.edge_path_to_video_path(video_path)
114
  if self.model_type != ModelType.ControlNetCanny:
115
  controlnet = ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-canny")
116
- self.set_model(ModelType.ControlNetCanny, model_id="runwayml/stable-diffusion-v1-5", controlnet=controlnet)
117
- self.pipe.scheduler = DDIMScheduler.from_config(self.pipe.scheduler.config)
118
- self.pipe.unet.set_attn_processor(processor=self.controlnet_attn_proc)
119
- self.pipe.controlnet.set_attn_processor(processor=self.controlnet_attn_proc)
 
 
 
 
120
 
121
- # TODO: Check scheduler
122
  added_prompt = 'best quality, extremely detailed'
123
  negative_prompts = 'longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality'
124
 
125
- video, fps = utils.prepare_video(video_path, resolution, self.device, self.dtype, False, start_t=0, end_t=15)
126
- control = utils.pre_process_canny(video, low_threshold, high_threshold).to(self.device).to(self.dtype)
 
 
127
  f, _, h, w = video.shape
128
  self.generator.manual_seed(seed)
129
- latents = torch.randn((1, 4, h//8, w//8), dtype=self.dtype, device=self.device, generator=self.generator)
 
130
  latents = latents.repeat(f, 1, 1, 1)
131
  result = self.inference(image=control,
132
  prompt=prompt + ', ' + added_prompt,
@@ -141,35 +161,49 @@ class Model:
141
  seed=seed,
142
  output_type='numpy',
143
  split_to_chunks=True,
144
- chunk_size=8,
145
  )
146
- return utils.create_video(result, fps)
147
 
148
  def process_controlnet_pose(self,
149
  video_path,
150
  prompt,
 
 
151
  num_inference_steps=20,
152
  controlnet_conditioning_scale=1.0,
153
  guidance_scale=9.0,
154
  seed=42,
155
  eta=0.0,
156
- resolution=512):
 
 
157
  video_path = gradio_utils.motion_to_video_path(video_path)
158
  if self.model_type != ModelType.ControlNetPose:
159
  controlnet = ControlNetModel.from_pretrained("fusing/stable-diffusion-v1-5-controlnet-openpose")
160
  self.set_model(ModelType.ControlNetPose, model_id="runwayml/stable-diffusion-v1-5", controlnet=controlnet)
161
- self.pipe.scheduler = DDIMScheduler.from_config(self.pipe.scheduler.config)
162
- self.pipe.unet.set_attn_processor(processor=self.controlnet_attn_proc)
163
- self.pipe.controlnet.set_attn_processor(processor=self.controlnet_attn_proc)
 
 
 
 
 
 
 
164
 
165
  added_prompt = 'best quality, extremely detailed, HD, ultra-realistic, 8K, HQ, masterpiece, trending on artstation, art, smooth'
166
  negative_prompts = 'longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer difits, cropped, worst quality, low quality, deformed body, bloated, ugly, unrealistic'
167
 
168
- video, fps = utils.prepare_video(video_path, resolution, self.device, self.dtype, False, output_fps=4)
169
- control = utils.pre_process_pose(video, apply_pose_detect=False).to(self.device).to(self.dtype)
 
 
170
  f, _, h, w = video.shape
171
  self.generator.manual_seed(seed)
172
- latents = torch.randn((1, 4, h//8, w//8), dtype=self.dtype, device=self.device, generator=self.generator)
 
173
  latents = latents.repeat(f, 1, 1, 1)
174
  result = self.inference(image=control,
175
  prompt=prompt + ', ' + added_prompt,
@@ -184,15 +218,16 @@ class Model:
184
  seed=seed,
185
  output_type='numpy',
186
  split_to_chunks=True,
187
- chunk_size=8,
188
  )
189
- return utils.create_gif(result, fps)
190
- # return utils.create_video(result, fps)
191
 
192
  def process_controlnet_canny_db(self,
193
  db_path,
194
  video_path,
195
  prompt,
 
 
196
  num_inference_steps=20,
197
  controlnet_conditioning_scale=1.0,
198
  guidance_scale=9.0,
@@ -200,26 +235,36 @@ class Model:
200
  eta=0.0,
201
  low_threshold=100,
202
  high_threshold=200,
203
- resolution=512):
 
 
204
  db_path = gradio_utils.get_model_from_db_selection(db_path)
205
  video_path = gradio_utils.get_video_from_canny_selection(video_path)
206
  # Load db and controlnet weights
207
  if 'db_path' not in self.states or db_path != self.states['db_path']:
208
  controlnet = ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-canny")
209
  self.set_model(ModelType.ControlNetCannyDB, model_id=db_path, controlnet=controlnet)
210
- self.pipe.scheduler = DDIMScheduler.from_config(self.pipe.scheduler.config)
211
- self.pipe.unet.set_attn_processor(processor=self.controlnet_attn_proc)
212
- self.pipe.controlnet.set_attn_processor(processor=self.controlnet_attn_proc)
213
  self.states['db_path'] = db_path
214
 
 
 
 
 
 
 
215
  added_prompt = 'best quality, extremely detailed'
216
  negative_prompts = 'longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality'
217
 
218
- video, fps = utils.prepare_video(video_path, resolution, self.device, self.dtype, False)
219
- control = utils.pre_process_canny(video, low_threshold, high_threshold).to(self.device).to(self.dtype)
 
 
220
  f, _, h, w = video.shape
221
  self.generator.manual_seed(seed)
222
- latents = torch.randn((1, 4, h//8, w//8), dtype=self.dtype, device=self.device, generator=self.generator)
 
223
  latents = latents.repeat(f, 1, 1, 1)
224
  result = self.inference(image=control,
225
  prompt=prompt + ', ' + added_prompt,
@@ -234,57 +279,91 @@ class Model:
234
  seed=seed,
235
  output_type='numpy',
236
  split_to_chunks=True,
237
- chunk_size=8,
238
  )
239
- return utils.create_gif(result, fps)
240
 
241
- def process_pix2pix(self, video, prompt, resolution=512, seed=0, start_t=0, end_t=-1, out_fps=-1):
242
- end_t = start_t+15
 
 
 
 
 
 
 
 
 
 
 
243
  if self.model_type != ModelType.Pix2Pix_Video:
244
- self.set_model(ModelType.Pix2Pix_Video, model_id="timbrooks/instruct-pix2pix")
245
- self.pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(self.pipe.scheduler.config)
246
- self.pipe.unet.set_attn_processor(processor=self.pix2pix_attn_proc)
247
- video, fps = utils.prepare_video(video, resolution, self.device, self.dtype, True, start_t, end_t, out_fps)
 
 
 
 
 
248
  self.generator.manual_seed(seed)
249
  result = self.inference(image=video,
250
  prompt=prompt,
251
  seed=seed,
252
  output_type='numpy',
253
  num_inference_steps=50,
254
- image_guidance_scale=1.5,
255
  split_to_chunks=True,
256
- chunk_size=8,
257
  )
258
- return utils.create_video(result, fps)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
259
 
260
- def process_text2video(self, prompt, motion_field_strength_x=12,motion_field_strength_y=12, n_prompt="", resolution=512, seed=24, num_frames=8, fps=2, t0=881, t1=941,
261
- use_cf_attn=True, use_motion_field=True,
262
- smooth_bg=False, smooth_bg_strength=0.4 ):
263
-
264
  if self.model_type != ModelType.Text2Video:
265
- unet = UNet2DConditionModel.from_pretrained('runwayml/stable-diffusion-v1-5', subfolder="unet")
266
- self.set_model(ModelType.Text2Video, model_id="runwayml/stable-diffusion-v1-5", unet=unet)
267
- self.pipe.scheduler = DDIMScheduler.from_config(self.pipe.scheduler.config)
 
268
  if use_cf_attn:
269
- self.pipe.unet.set_attn_processor(processor=self.text2video_attn_proc)
 
270
  self.generator.manual_seed(seed)
271
 
272
-
273
  added_prompt = "high quality, HD, 8K, trending on artstation, high focus, dramatic lighting"
274
  negative_prompts = 'longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer difits, cropped, worst quality, low quality, deformed body, bloated, ugly, unrealistic'
275
 
276
  prompt = prompt.rstrip()
277
- if len(prompt) > 0 and (prompt[-1] == "," or prompt[-1] == "."):
278
  prompt = prompt.rstrip()[:-1]
279
  prompt = prompt.rstrip()
280
  prompt = prompt + ", "+added_prompt
281
- if len(n_prompt)>0:
282
- negative_prompt = [n_prompt]
283
  else:
284
  negative_prompt = None
285
 
286
- result = self.inference(prompt=[prompt],
287
- video_length=num_frames,
288
  height=resolution,
289
  width=resolution,
290
  num_inference_steps=50,
@@ -299,6 +378,9 @@ class Model:
299
  smooth_bg_strength=smooth_bg_strength,
300
  seed=seed,
301
  output_type='numpy',
302
- negative_prompt = negative_prompt,
 
 
 
303
  )
304
- return utils.create_video(result, fps)
 
11
  import utils
12
  import gradio_utils
13
 
14
+ # decord.bridge.set_bridge('torch')
15
 
16
 
17
  class ModelType(Enum):
 
55
  def inference_chunk(self, frame_ids, **kwargs):
56
  if self.pipe is None:
57
  return
58
+
59
  prompt = np.array(kwargs.pop('prompt'))
60
  negative_prompt = np.array(kwargs.pop('negative_prompt', ''))
61
  latents = None
62
  if 'latents' in kwargs:
63
  latents = kwargs.pop('latents')[frame_ids]
64
+ if 'image' in kwargs:
65
+ kwargs['image'] = kwargs['image'][frame_ids]
66
+ if 'video_length' in kwargs:
67
+ kwargs['video_length'] = len(frame_ids)
68
+ if self.model_type == ModelType.Text2Video:
69
+ kwargs["frame_ids"] = frame_ids
70
+ return self.pipe(prompt=prompt[frame_ids].tolist(),
71
  negative_prompt=negative_prompt[frame_ids].tolist(),
72
  latents=latents,
73
  generator=self.generator,
 
77
  if self.pipe is None:
78
  return
79
  seed = kwargs.pop('seed', 0)
80
+ if seed < 0:
81
+ seed = self.generator.seed()
82
  kwargs.pop('generator', '')
83
+
84
+ if 'image' in kwargs:
85
+ f = kwargs['image'].shape[0]
86
+ else:
87
+ f = kwargs['video_length']
88
+
89
+ assert 'prompt' in kwargs
90
+ prompt = [kwargs.pop('prompt')] * f
91
+ negative_prompt = [kwargs.pop('negative_prompt', '')] * f
92
+
93
+ # Processing chunk-by-chunk
94
  if split_to_chunks:
 
 
 
 
 
 
95
  chunk_ids = np.arange(0, f, chunk_size - 1)
96
  result = []
97
  for i in range(len(chunk_ids)):
 
101
  self.generator.manual_seed(seed)
102
  print(f'Processing chunk {i + 1} / {len(chunk_ids)}')
103
  result.append(self.inference_chunk(frame_ids=frame_ids,
104
+ prompt=prompt,
105
+ negative_prompt=negative_prompt,
 
106
  **kwargs).images[1:])
107
  result = np.concatenate(result)
108
  return result
109
  else:
110
+ return self.pipe(prompt=prompt, negative_prompt=negative_prompt, generator=self.generator, **kwargs).images
111
 
112
  def process_controlnet_canny(self,
113
  video_path,
114
  prompt,
115
+ chunk_size=8,
116
+ watermark=None,
117
  num_inference_steps=20,
118
  controlnet_conditioning_scale=1.0,
119
  guidance_scale=9.0,
 
121
  eta=0.0,
122
  low_threshold=100,
123
  high_threshold=200,
124
+ resolution=512,
125
+ use_cf_attn=True,
126
+ save_path=None):
127
  video_path = gradio_utils.edge_path_to_video_path(video_path)
128
  if self.model_type != ModelType.ControlNetCanny:
129
  controlnet = ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-canny")
130
+ self.set_model(ModelType.ControlNetCanny,model_id="runwayml/stable-diffusion-v1-5", controlnet=controlnet)
131
+ self.pipe.scheduler = DDIMScheduler.from_config(
132
+ self.pipe.scheduler.config)
133
+ if use_cf_attn:
134
+ self.pipe.unet.set_attn_processor(
135
+ processor=self.controlnet_attn_proc)
136
+ self.pipe.controlnet.set_attn_processor(
137
+ processor=self.controlnet_attn_proc)
138
 
 
139
  added_prompt = 'best quality, extremely detailed'
140
  negative_prompts = 'longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality'
141
 
142
+ video, fps = utils.prepare_video(
143
+ video_path, resolution, self.device, self.dtype, False)
144
+ control = utils.pre_process_canny(
145
+ video, low_threshold, high_threshold).to(self.device).to(self.dtype)
146
  f, _, h, w = video.shape
147
  self.generator.manual_seed(seed)
148
+ latents = torch.randn((1, 4, h//8, w//8), dtype=self.dtype,
149
+ device=self.device, generator=self.generator)
150
  latents = latents.repeat(f, 1, 1, 1)
151
  result = self.inference(image=control,
152
  prompt=prompt + ', ' + added_prompt,
 
161
  seed=seed,
162
  output_type='numpy',
163
  split_to_chunks=True,
164
+ chunk_size=chunk_size,
165
  )
166
+ return utils.create_video(result, fps, path=save_path, watermark=gradio_utils.logo_name_to_path(watermark))
167
 
168
  def process_controlnet_pose(self,
169
  video_path,
170
  prompt,
171
+ chunk_size=8,
172
+ watermark=None,
173
  num_inference_steps=20,
174
  controlnet_conditioning_scale=1.0,
175
  guidance_scale=9.0,
176
  seed=42,
177
  eta=0.0,
178
+ resolution=512,
179
+ use_cf_attn=True,
180
+ save_path=None):
181
  video_path = gradio_utils.motion_to_video_path(video_path)
182
  if self.model_type != ModelType.ControlNetPose:
183
  controlnet = ControlNetModel.from_pretrained("fusing/stable-diffusion-v1-5-controlnet-openpose")
184
  self.set_model(ModelType.ControlNetPose, model_id="runwayml/stable-diffusion-v1-5", controlnet=controlnet)
185
+ self.pipe.scheduler = DDIMScheduler.from_config(
186
+ self.pipe.scheduler.config)
187
+ if use_cf_attn:
188
+ self.pipe.unet.set_attn_processor(
189
+ processor=self.controlnet_attn_proc)
190
+ self.pipe.controlnet.set_attn_processor(
191
+ processor=self.controlnet_attn_proc)
192
+
193
+ video_path = gradio_utils.motion_to_video_path(
194
+ video_path) if 'Motion' in video_path else video_path
195
 
196
  added_prompt = 'best quality, extremely detailed, HD, ultra-realistic, 8K, HQ, masterpiece, trending on artstation, art, smooth'
197
  negative_prompts = 'longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer difits, cropped, worst quality, low quality, deformed body, bloated, ugly, unrealistic'
198
 
199
+ video, fps = utils.prepare_video(
200
+ video_path, resolution, self.device, self.dtype, False, output_fps=4)
201
+ control = utils.pre_process_pose(
202
+ video, apply_pose_detect=False).to(self.device).to(self.dtype)
203
  f, _, h, w = video.shape
204
  self.generator.manual_seed(seed)
205
+ latents = torch.randn((1, 4, h//8, w//8), dtype=self.dtype,
206
+ device=self.device, generator=self.generator)
207
  latents = latents.repeat(f, 1, 1, 1)
208
  result = self.inference(image=control,
209
  prompt=prompt + ', ' + added_prompt,
 
218
  seed=seed,
219
  output_type='numpy',
220
  split_to_chunks=True,
221
+ chunk_size=chunk_size,
222
  )
223
+ return utils.create_gif(result, fps, path=save_path, watermark=gradio_utils.logo_name_to_path(watermark))
 
224
 
225
  def process_controlnet_canny_db(self,
226
  db_path,
227
  video_path,
228
  prompt,
229
+ chunk_size=8,
230
+ watermark=None,
231
  num_inference_steps=20,
232
  controlnet_conditioning_scale=1.0,
233
  guidance_scale=9.0,
 
235
  eta=0.0,
236
  low_threshold=100,
237
  high_threshold=200,
238
+ resolution=512,
239
+ use_cf_attn=True,
240
+ save_path=None):
241
  db_path = gradio_utils.get_model_from_db_selection(db_path)
242
  video_path = gradio_utils.get_video_from_canny_selection(video_path)
243
  # Load db and controlnet weights
244
  if 'db_path' not in self.states or db_path != self.states['db_path']:
245
  controlnet = ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-canny")
246
  self.set_model(ModelType.ControlNetCannyDB, model_id=db_path, controlnet=controlnet)
247
+ self.pipe.scheduler = DDIMScheduler.from_config(
248
+ self.pipe.scheduler.config)
 
249
  self.states['db_path'] = db_path
250
 
251
+ if use_cf_attn:
252
+ self.pipe.unet.set_attn_processor(
253
+ processor=self.controlnet_attn_proc)
254
+ self.pipe.controlnet.set_attn_processor(
255
+ processor=self.controlnet_attn_proc)
256
+
257
  added_prompt = 'best quality, extremely detailed'
258
  negative_prompts = 'longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality'
259
 
260
+ video, fps = utils.prepare_video(
261
+ video_path, resolution, self.device, self.dtype, False)
262
+ control = utils.pre_process_canny(
263
+ video, low_threshold, high_threshold).to(self.device).to(self.dtype)
264
  f, _, h, w = video.shape
265
  self.generator.manual_seed(seed)
266
+ latents = torch.randn((1, 4, h//8, w//8), dtype=self.dtype,
267
+ device=self.device, generator=self.generator)
268
  latents = latents.repeat(f, 1, 1, 1)
269
  result = self.inference(image=control,
270
  prompt=prompt + ', ' + added_prompt,
 
279
  seed=seed,
280
  output_type='numpy',
281
  split_to_chunks=True,
282
+ chunk_size=chunk_size,
283
  )
284
+ return utils.create_gif(result, fps, path=save_path, watermark=gradio_utils.logo_name_to_path(watermark))
285
 
286
+ def process_pix2pix(self,
287
+ video,
288
+ prompt,
289
+ resolution=512,
290
+ seed=0,
291
+ image_guidance_scale=1.0,
292
+ start_t=0,
293
+ end_t=-1,
294
+ out_fps=-1,
295
+ chunk_size=8,
296
+ watermark=None,
297
+ use_cf_attn=True,
298
+ save_path=None,):
299
  if self.model_type != ModelType.Pix2Pix_Video:
300
+ self.set_model(ModelType.Pix2Pix_Video,
301
+ model_id="timbrooks/instruct-pix2pix")
302
+ self.pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(
303
+ self.pipe.scheduler.config)
304
+ if use_cf_attn:
305
+ self.pipe.unet.set_attn_processor(
306
+ processor=self.pix2pix_attn_proc)
307
+ video, fps = utils.prepare_video(
308
+ video, resolution, self.device, self.dtype, True, start_t, end_t, out_fps)
309
  self.generator.manual_seed(seed)
310
  result = self.inference(image=video,
311
  prompt=prompt,
312
  seed=seed,
313
  output_type='numpy',
314
  num_inference_steps=50,
315
+ image_guidance_scale=image_guidance_scale,
316
  split_to_chunks=True,
317
+ chunk_size=chunk_size,
318
  )
319
+ return utils.create_video(result, fps, path=save_path, watermark=gradio_utils.logo_name_to_path(watermark))
320
+
321
+ def process_text2video(self,
322
+ prompt,
323
+ model_name,
324
+ motion_field_strength_x=12,
325
+ motion_field_strength_y=12,
326
+ t0=44,
327
+ t1=47,
328
+ n_prompt="",
329
+ chunk_size=8,
330
+ video_length=8,
331
+ watermark=None,
332
+ inject_noise_to_warp=False,
333
+ resolution=512,
334
+ seed=-1,
335
+ fps=2,
336
+ use_cf_attn=True,
337
+ use_motion_field=True,
338
+ smooth_bg=False,
339
+ smooth_bg_strength=0.4,
340
+ path=None):
341
 
 
 
 
 
342
  if self.model_type != ModelType.Text2Video:
343
+ unet = UNet2DConditionModel.from_pretrained(model_name, subfolder="unet")
344
+ self.set_model(ModelType.Text2Video, model_id=model_name, unet=unet)
345
+ self.pipe.scheduler = DDIMScheduler.from_config(
346
+ self.pipe.scheduler.config)
347
  if use_cf_attn:
348
+ self.pipe.unet.set_attn_processor(
349
+ processor=self.text2video_attn_proc)
350
  self.generator.manual_seed(seed)
351
 
 
352
  added_prompt = "high quality, HD, 8K, trending on artstation, high focus, dramatic lighting"
353
  negative_prompts = 'longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer difits, cropped, worst quality, low quality, deformed body, bloated, ugly, unrealistic'
354
 
355
  prompt = prompt.rstrip()
356
+ if len(prompt) > 0 and (prompt[-1] == "," or prompt[-1] == "."):
357
  prompt = prompt.rstrip()[:-1]
358
  prompt = prompt.rstrip()
359
  prompt = prompt + ", "+added_prompt
360
+ if len(n_prompt) > 0:
361
+ negative_prompt = n_prompt
362
  else:
363
  negative_prompt = None
364
 
365
+ result = self.inference(prompt=prompt,
366
+ video_length=video_length,
367
  height=resolution,
368
  width=resolution,
369
  num_inference_steps=50,
 
378
  smooth_bg_strength=smooth_bg_strength,
379
  seed=seed,
380
  output_type='numpy',
381
+ negative_prompt=negative_prompt,
382
+ inject_noise_to_warp=inject_noise_to_warp,
383
+ split_to_chunks=True,
384
+ chunk_size=chunk_size,
385
  )
386
+ return utils.create_video(result, fps, path=path, watermark=gradio_utils.logo_name_to_path(watermark))
text_to_video/text_to_video_generator.py DELETED
@@ -1,79 +0,0 @@
1
- from text_to_video.tuneavideo.pipelines.pipeline_text_to_video import TuneAVideoPipeline
2
- from text_to_video.tuneavideo.models.unet import UNet3DConditionModel
3
- import torch
4
- from diffusers import AutoencoderKL, DDIMScheduler
5
- from transformers import CLIPTextModel, CLIPTokenizer
6
-
7
-
8
- class TextToVideo():
9
-
10
-
11
- def __init__(self,sd_path = None,motion_field_strength = 12, video_length = 8,t0 = 881, t1=941,use_cf_attn=True,use_motion_field=True) -> None:
12
- g = torch.Generator(device='cuda')
13
- g.manual_seed(22)
14
- self.g = g
15
-
16
- assert sd_path is not None
17
-
18
- print(f"Loading model SD-Net model file from {sd_path}")
19
-
20
- self.dtype = torch.float16
21
- noise_scheduler = DDIMScheduler.from_pretrained(
22
- sd_path, subfolder="scheduler")
23
- tokenizer = CLIPTokenizer.from_pretrained(
24
- sd_path, subfolder="tokenizer")
25
- text_encoder = CLIPTextModel.from_pretrained(
26
- sd_path, subfolder="text_encoder")
27
- vae = AutoencoderKL.from_pretrained(sd_path, subfolder="vae")
28
-
29
-
30
- unet = UNet3DConditionModel.from_pretrained_2d(
31
- sd_path, subfolder="unet", use_cf_attn=use_cf_attn)
32
- self.pipe = TuneAVideoPipeline(
33
- vae=vae, text_encoder=text_encoder, tokenizer=tokenizer, unet=unet,
34
- scheduler=DDIMScheduler.from_pretrained(
35
- sd_path, subfolder="scheduler")
36
- ).to('cuda').to(self.dtype)
37
-
38
- noise_scheduler.set_timesteps(50, device='cuda')
39
-
40
- # t0 parameter (DDIM backward from noise until t0)
41
- self.t0 = t0
42
-
43
-
44
- # from t0 apply DDPM forward until t1
45
- self.t1 = t1
46
-
47
- self.use_foreground_motion_field = False # apply motion field on forground object (not used)
48
-
49
- # strength of motion field (delta_x = delta_y in Sect 3.3.1)
50
- self.motion_field_strength = motion_field_strength
51
- self.use_motion_field = use_motion_field # apply general motion field
52
- self.smooth_bg = False # temporally smooth background
53
- self.smooth_bg_strength = 0.4 # alpha = (1-self.smooth_bg_strength) in Eq (9)
54
-
55
-
56
- self.video_length = video_length
57
-
58
- def inference(self, prompt):
59
-
60
- prompt_compute = [prompt]
61
- xT = torch.randn((1, 4, 1, 64, 64), dtype=self.dtype, device="cuda")
62
- result = self.pipe(prompt_compute,
63
- video_length=self.video_length,
64
- height=512,
65
- width=512,
66
- num_inference_steps=50,
67
- guidance_scale=7.5,
68
- guidance_stop_step=1.0,
69
- t0=self.t0,
70
- t1=self.t1,
71
- xT=xT,
72
- use_foreground_motion_field=self.use_foreground_motion_field,
73
- motion_field_strength=self.motion_field_strength,
74
- use_motion_field=self.use_motion_field,
75
- smooth_bg=self.smooth_bg,
76
- smooth_bg_strength=self.smooth_bg_strength,
77
- generator=self.g)
78
-
79
- return result.videos[0]
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
text_to_video/text_to_video_pipeline.py CHANGED
@@ -11,22 +11,27 @@ from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
11
  from diffusers.models import AutoencoderKL, UNet2DConditionModel
12
  from diffusers.schedulers import KarrasDiffusionSchedulers
13
  from diffusers.pipelines.stable_diffusion import StableDiffusionSafetyChecker
 
 
 
 
14
 
15
  @dataclass
16
  class TextToVideoPipelineOutput(BaseOutput):
17
- videos: Union[torch.Tensor, np.ndarray]
18
- code: Union[torch.Tensor, np.ndarray]
19
-
 
20
 
21
 
22
  def coords_grid(batch, ht, wd, device):
23
  # Adapted from https://github.com/princeton-vl/RAFT/blob/master/core/utils/utils.py
24
- coords = torch.meshgrid(torch.arange(ht, device=device), torch.arange(wd, device=device))
 
25
  coords = torch.stack(coords[::-1], dim=0).float()
26
  return coords[None].repeat(batch, 1, 1, 1)
27
 
28
 
29
-
30
  class TextToVideoPipeline(StableDiffusionPipeline):
31
  def __init__(
32
  self,
@@ -38,14 +43,13 @@ class TextToVideoPipeline(StableDiffusionPipeline):
38
  safety_checker: StableDiffusionSafetyChecker,
39
  feature_extractor: CLIPFeatureExtractor,
40
  requires_safety_checker: bool = True,
41
- ):
42
- #super().__init__(*args,**kwargs)
43
- super().__init__(vae,text_encoder,tokenizer,unet,scheduler,safety_checker,feature_extractor,requires_safety_checker)
44
-
45
 
46
  def DDPM_forward(self, x0, t0, tMax, generator, device, shape, text_embeddings):
47
  rand_device = "cpu" if device.type == "mps" else device
48
-
49
  if x0 is None:
50
  return torch.randn(shape, generator=generator, device=rand_device, dtype=text_embeddings.dtype).to(device)
51
  else:
@@ -55,7 +59,6 @@ class TextToVideoPipeline(StableDiffusionPipeline):
55
  torch.sqrt(1-alpha_vec) * eps
56
  return xt
57
 
58
-
59
  def prepare_latents(self, batch_size, num_channels_latents, video_length, height, width, dtype, device, generator, latents=None):
60
  shape = (batch_size, num_channels_latents, video_length, height //
61
  self.vae_scale_factor, width // self.vae_scale_factor)
@@ -86,9 +89,7 @@ class TextToVideoPipeline(StableDiffusionPipeline):
86
  latents = latents * self.scheduler.init_noise_sigma
87
  return latents
88
 
89
-
90
-
91
- def warp_latents(self, latents, reference_flow):
92
  _, _, H, W = reference_flow.size()
93
  b, c, f, h, w = latents.size()
94
  coords0 = coords_grid(f, H, W, device=latents.device).to(latents.dtype)
@@ -101,19 +102,20 @@ class TextToVideoPipeline(StableDiffusionPipeline):
101
  latents_0 = latents[:, :, 0]
102
  latents_0 = latents_0.repeat(f, 1, 1, 1)
103
  warped = grid_sample(latents_0, coords_t0,
104
- mode='nearest', padding_mode='reflection')
105
  warped = rearrange(warped, '(b f) c h w -> b c f h w', f=f)
106
  return warped
107
 
108
- def warp_latents_independently(self, latents, reference_flow):
109
  _, _, H, W = reference_flow.size()
110
- b, c, f, h, w = latents.size()
111
  assert b == 1
112
  coords0 = coords_grid(f, H, W, device=latents.device).to(latents.dtype)
113
- coords_t0 = coords0 + reference_flow
114
 
 
115
  coords_t0[:, 0] /= W
116
  coords_t0[:, 1] /= H
 
117
  coords_t0 = coords_t0 * 2.0 - 1.0
118
 
119
  coords_t0 = T.Resize((h, w))(coords_t0)
@@ -121,23 +123,32 @@ class TextToVideoPipeline(StableDiffusionPipeline):
121
  coords_t0 = rearrange(coords_t0, 'f c h w -> f h w c')
122
 
123
  latents_0 = rearrange(latents[0], 'c f h w -> f c h w')
124
-
125
  warped = grid_sample(latents_0, coords_t0,
126
  mode='nearest', padding_mode='reflection')
 
 
 
 
 
 
 
 
 
 
127
  warped = rearrange(warped, '(b f) c h w -> b c f h w', f=f)
128
  return warped
129
 
130
- def DDIM_backward(self, num_inference_steps, timesteps, skip_t, t0, t1, do_classifier_free_guidance, null_embs, text_embeddings, latents_local, latents_dtype, guidance_scale, guidance_stop_step, callback, callback_steps, extra_step_kwargs, num_warmup_steps):
 
131
  entered = False
132
-
133
  f = latents_local.shape[2]
134
- latents_local = rearrange(latents_local,"b c f w h -> (b f) c w h")
135
-
 
136
  latents = latents_local.detach().clone()
137
  x_t0_1 = None
138
  x_t1_1 = None
139
-
140
-
141
 
142
  with self.progress_bar(total=num_inference_steps) as progress_bar:
143
  for i, t in enumerate(timesteps):
@@ -160,7 +171,8 @@ class TextToVideoPipeline(StableDiffusionPipeline):
160
  with torch.no_grad():
161
  if null_embs is not None:
162
  text_embeddings[0] = null_embs[i][0]
163
- te = torch.cat([repeat(text_embeddings[0,:,:], "c k -> f c k",f=f),repeat(text_embeddings[1,:,:], "c k -> f c k",f=f)])
 
164
  noise_pred = self.unet(
165
  latent_model_input, t, encoder_hidden_states=te).sample.to(dtype=latents_dtype)
166
 
@@ -191,17 +203,14 @@ class TextToVideoPipeline(StableDiffusionPipeline):
191
  if callback is not None and i % callback_steps == 0:
192
  callback(i, t, latents)
193
 
194
-
195
- latents = rearrange(latents,"(b f) c w h -> b c f w h",f = f)
196
-
197
-
198
-
199
  res = {"x0": latents.detach().clone()}
200
  if x_t0_1 is not None:
201
- x_t0_1 = rearrange(x_t0_1,"(b f) c w h -> b c f w h",f = f)
202
  res["x_t0_1"] = x_t0_1.detach().clone()
203
  if x_t1_1 is not None:
204
- x_t1_1 = rearrange(x_t1_1,"(b f) c w h -> b c f w h",f = f)
205
  res["x_t1_1"] = x_t1_1.detach().clone()
206
  return res
207
 
@@ -213,9 +222,28 @@ class TextToVideoPipeline(StableDiffusionPipeline):
213
  video = rearrange(video, "(b f) c h w -> b c f h w", f=video_length)
214
  video = (video / 2 + 0.5).clamp(0, 1)
215
  # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
 
216
  return video
217
 
 
 
 
 
 
 
 
 
 
 
218
 
 
 
 
 
 
 
 
 
219
 
220
  @torch.no_grad()
221
  def __call__(
@@ -234,22 +262,46 @@ class TextToVideoPipeline(StableDiffusionPipeline):
234
  List[torch.Generator]]] = None,
235
  xT: Optional[torch.FloatTensor] = None,
236
  null_embs: Optional[torch.FloatTensor] = None,
237
- #motion_field_strength: float = 12,
238
  motion_field_strength_x: float = 12,
239
- motion_field_strength_y: float = 12,
240
  output_type: Optional[str] = "tensor",
241
  return_dict: bool = True,
242
  callback: Optional[Callable[[
243
  int, int, torch.FloatTensor], None]] = None,
244
  callback_steps: Optional[int] = 1,
245
  use_motion_field: bool = True,
246
- smooth_bg: bool = True,
247
  smooth_bg_strength: float = 0.4,
 
 
 
248
  **kwargs,
249
  ):
250
- print(motion_field_strength_x,motion_field_strength_y)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
251
  print(f" Use: Motion field = {use_motion_field}")
252
  print(f" Use: Background smoothing = {smooth_bg}")
 
253
  # Default height and width to unet
254
  height = height or self.unet.config.sample_size * self.vae_scale_factor
255
  width = width or self.unet.config.sample_size * self.vae_scale_factor
@@ -269,11 +321,11 @@ class TextToVideoPipeline(StableDiffusionPipeline):
269
  text_embeddings = self._encode_prompt(
270
  prompt, device, num_videos_per_prompt, do_classifier_free_guidance, negative_prompt
271
  )
272
-
273
  # Prepare timesteps
274
  self.scheduler.set_timesteps(num_inference_steps, device=device)
275
  timesteps = self.scheduler.timesteps
276
-
277
  # print(f" Latent shape = {latents.shape}")
278
 
279
  # Prepare latent variables
@@ -282,7 +334,7 @@ class TextToVideoPipeline(StableDiffusionPipeline):
282
  xT = self.prepare_latents(
283
  batch_size * num_videos_per_prompt,
284
  num_channels_latents,
285
- video_length,
286
  height,
287
  width,
288
  text_embeddings.dtype,
@@ -309,27 +361,43 @@ class TextToVideoPipeline(StableDiffusionPipeline):
309
  None,
310
  )
311
  xT = torch.cat([xT, xT_missing], dim=2)
312
-
313
 
314
  xInit = xT.clone()
315
- t0 = kwargs["t0"]
316
- t1 = kwargs["t1"]
 
 
 
 
 
 
 
 
317
  x_t1_1 = None
318
 
319
-
320
  # Prepare extra step kwargs.
321
  extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
322
  # Denoising loop
323
  num_warmup_steps = len(timesteps) - \
324
  num_inference_steps * self.scheduler.order
325
-
326
 
 
 
 
 
 
 
 
 
 
 
327
 
328
  ddim_res = self.DDIM_backward(num_inference_steps=num_inference_steps, timesteps=timesteps, skip_t=1000, t0=t0, t1=t1, do_classifier_free_guidance=do_classifier_free_guidance,
329
- null_embs=null_embs, text_embeddings=text_embeddings, latents_local=xT, latents_dtype=dtype, guidance_scale=guidance_scale, guidance_stop_step=guidance_stop_step, callback=callback, callback_steps=callback_steps, extra_step_kwargs=extra_step_kwargs, num_warmup_steps=num_warmup_steps)
330
-
 
331
  x0 = ddim_res["x0"].detach()
332
-
333
  if "x_t0_1" in ddim_res:
334
  x_t0_1 = ddim_res["x_t0_1"].detach()
335
  if "x_t1_1" in ddim_res:
@@ -337,25 +405,36 @@ class TextToVideoPipeline(StableDiffusionPipeline):
337
  del ddim_res
338
  del xT
339
 
340
- if use_motion_field:
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
341
  del x0
342
- shape = (batch_size, num_channels_latents, 1, height //
343
- self.vae_scale_factor, width // self.vae_scale_factor)
344
-
345
-
346
- x_t0_k = x_t0_1[:, :, :1, :, :].repeat(1, 1, video_length-1, 1, 1)
347
 
348
-
349
- reference_flow = torch.zeros(
350
- (video_length-1, 2, 512, 512), device=x_t0_1.device, dtype=x_t0_1.dtype)
351
- for fr_idx in range(video_length-1):
352
- #reference_flow[fr_idx, :, :, :] = motion_field_strength*(fr_idx+1)
353
- reference_flow[fr_idx, 0, :, :] = motion_field_strength_x*(fr_idx+1)
354
- reference_flow[fr_idx, 1, :, :] = motion_field_strength_y*(fr_idx+1)
355
 
356
- for idx, latent in enumerate(x_t0_k):
357
- x_t0_k[idx] = self.warp_latents_independently(
358
- latent[None], reference_flow)
359
 
360
  # assuming t0=t1=1000, if t0 = 1000
361
  if t1 > t0:
@@ -370,16 +449,21 @@ class TextToVideoPipeline(StableDiffusionPipeline):
370
  x_t1 = torch.cat([x_t1_1, x_t1_k], dim=2).clone().detach()
371
 
372
  ddim_res = self.DDIM_backward(num_inference_steps=num_inference_steps, timesteps=timesteps, skip_t=t1, t0=-1, t1=-1, do_classifier_free_guidance=do_classifier_free_guidance,
373
- null_embs=null_embs, text_embeddings=text_embeddings, latents_local=x_t1, latents_dtype=dtype, guidance_scale=guidance_scale, guidance_stop_step=guidance_stop_step, callback=callback, callback_steps=callback_steps, extra_step_kwargs=extra_step_kwargs, num_warmup_steps=num_warmup_steps)
 
374
 
375
  x0 = ddim_res["x0"].detach()
376
  del ddim_res
 
 
 
 
377
  else:
378
  x_t1 = x_t1_1.clone()
379
- x_t1_1 = x_t1_1[:,:,:1,:,:].clone()
380
- x_t1_k = x_t1_1[:,:,1:,:,:].clone()
381
  x_t0_k = x_t0_1[:, :, 1:, :, :].clone()
382
- x_t0_1 = x_t0_1[:,:,:1,:,:].clone()
383
 
384
  # smooth background
385
  if smooth_bg:
@@ -401,12 +485,10 @@ class TextToVideoPipeline(StableDiffusionPipeline):
401
  mask = dilation(mask[None].to(x0.device), kernel)[0]
402
  M_FG[batch_idx, frame_idx, :, :] = mask
403
 
404
-
405
  x_t1_1_fg_masked = x_t1_1 * \
406
  (1 - repeat(M_FG[:, 0, :, :],
407
  "b w h -> b c 1 w h", c=x_t1_1.shape[1]))
408
 
409
-
410
  x_t1_1_fg_masked_moved = []
411
  for batch_idx, x_t1_1_fg_masked_b in enumerate(x_t1_1_fg_masked):
412
  x_t1_fg_masked_b = x_t1_1_fg_masked_b.clone()
@@ -416,7 +498,7 @@ class TextToVideoPipeline(StableDiffusionPipeline):
416
  if use_motion_field:
417
  x_t1_fg_masked_b = x_t1_fg_masked_b[None]
418
  x_t1_fg_masked_b = self.warp_latents_independently(
419
- x_t1_fg_masked_b, reference_flow)
420
  else:
421
  x_t1_fg_masked_b = x_t1_fg_masked_b[None]
422
 
@@ -432,9 +514,9 @@ class TextToVideoPipeline(StableDiffusionPipeline):
432
  for batch_idx, m_fg_1_b in enumerate(M_FG_1):
433
  m_fg_1_b = m_fg_1_b[None, None]
434
  m_fg_b = m_fg_1_b.repeat(1, 1, video_length-1, 1, 1)
435
- if use_motion_field:
436
  m_fg_b = self.warp_latents_independently(
437
- m_fg_b.clone(), reference_flow)
438
  M_FG_warped.append(
439
  torch.cat([m_fg_1_b[:1, 0], m_fg_b[:1, 0]], dim=1))
440
 
@@ -445,45 +527,40 @@ class TextToVideoPipeline(StableDiffusionPipeline):
445
  M_BG = (1-M_FG) * (1 - M_FG_warped)
446
  M_BG = repeat(M_BG, "b f h w -> b c f h w", c=channels)
447
  a_convex = smooth_bg_strength
448
-
449
- x_t1_blending = (1-M_BG) * x_t1 + M_BG * (a_convex *
450
- x_t1 + (1-a_convex) * x_t1_1_fg_masked_moved)
451
 
452
- '''
453
- x_t1_blending = self.DDPM_forward(
454
- x0=x_t1_blending, t0=t1, tMax=961, device=device, shape=shape, text_embeddings=text_embeddings, generator=generator)
455
- t1 = 961
456
- '''
457
- latents = x_t1_blending
458
 
459
  ddim_res = self.DDIM_backward(num_inference_steps=num_inference_steps, timesteps=timesteps, skip_t=t1, t0=-1, t1=-1, do_classifier_free_guidance=do_classifier_free_guidance,
460
- null_embs=null_embs, text_embeddings=text_embeddings, latents_local=latents, latents_dtype=dtype, guidance_scale=guidance_scale, guidance_stop_step=guidance_stop_step, callback=callback, callback_steps=callback_steps, extra_step_kwargs=extra_step_kwargs, num_warmup_steps=num_warmup_steps)
 
461
  x0 = ddim_res["x0"].detach()
462
  del ddim_res
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
463
 
 
 
 
464
 
465
- # Post-processing
466
- video_list = []
467
- for latent in x0:
468
- tmp = latent[None]
469
- print("Frame spit shape", tmp.shape)
470
- frames = []
471
- for fr_split in range(tmp.shape[2]):
472
- print("frame decoding")
473
- frames.append(self.decode_latents(
474
- tmp[:, :, fr_split, None]).detach())
475
-
476
- video_list.append(torch.cat(frames, dim=2).cpu().float().numpy())
477
-
478
- # Convert to tensor
479
- videos = []
480
- if output_type == "tensor":
481
- for video in video_list:
482
- videos.append(torch.from_numpy(video))
483
- if output_type == 'numpy':
484
- for video in video_list:
485
- videos.append(rearrange(video, 'b c f h w -> (b f) h w c'))
486
  if not return_dict:
487
- return video
488
 
489
- return TextToVideoPipelineOutput(videos=videos, code=torch.split(xInit.detach().cpu(), 1, dim=0))
 
11
  from diffusers.models import AutoencoderKL, UNet2DConditionModel
12
  from diffusers.schedulers import KarrasDiffusionSchedulers
13
  from diffusers.pipelines.stable_diffusion import StableDiffusionSafetyChecker
14
+ import PIL
15
+ from PIL import Image
16
+ from kornia.morphology import dilation
17
+
18
 
19
  @dataclass
20
  class TextToVideoPipelineOutput(BaseOutput):
21
+ # videos: Union[torch.Tensor, np.ndarray]
22
+ # code: Union[torch.Tensor, np.ndarray]
23
+ images: Union[List[PIL.Image.Image], np.ndarray]
24
+ nsfw_content_detected: Optional[List[bool]]
25
 
26
 
27
  def coords_grid(batch, ht, wd, device):
28
  # Adapted from https://github.com/princeton-vl/RAFT/blob/master/core/utils/utils.py
29
+ coords = torch.meshgrid(torch.arange(
30
+ ht, device=device), torch.arange(wd, device=device))
31
  coords = torch.stack(coords[::-1], dim=0).float()
32
  return coords[None].repeat(batch, 1, 1, 1)
33
 
34
 
 
35
  class TextToVideoPipeline(StableDiffusionPipeline):
36
  def __init__(
37
  self,
 
43
  safety_checker: StableDiffusionSafetyChecker,
44
  feature_extractor: CLIPFeatureExtractor,
45
  requires_safety_checker: bool = True,
46
+ ):
47
+ super().__init__(vae, text_encoder, tokenizer, unet, scheduler,
48
+ safety_checker, feature_extractor, requires_safety_checker)
 
49
 
50
  def DDPM_forward(self, x0, t0, tMax, generator, device, shape, text_embeddings):
51
  rand_device = "cpu" if device.type == "mps" else device
52
+
53
  if x0 is None:
54
  return torch.randn(shape, generator=generator, device=rand_device, dtype=text_embeddings.dtype).to(device)
55
  else:
 
59
  torch.sqrt(1-alpha_vec) * eps
60
  return xt
61
 
 
62
  def prepare_latents(self, batch_size, num_channels_latents, video_length, height, width, dtype, device, generator, latents=None):
63
  shape = (batch_size, num_channels_latents, video_length, height //
64
  self.vae_scale_factor, width // self.vae_scale_factor)
 
89
  latents = latents * self.scheduler.init_noise_sigma
90
  return latents
91
 
92
+ def warp_latents_from_f0(self, latents, reference_flow):
 
 
93
  _, _, H, W = reference_flow.size()
94
  b, c, f, h, w = latents.size()
95
  coords0 = coords_grid(f, H, W, device=latents.device).to(latents.dtype)
 
102
  latents_0 = latents[:, :, 0]
103
  latents_0 = latents_0.repeat(f, 1, 1, 1)
104
  warped = grid_sample(latents_0, coords_t0,
105
+ mode='nearest', padding_mode='reflection')
106
  warped = rearrange(warped, '(b f) c h w -> b c f h w', f=f)
107
  return warped
108
 
109
+ def warp_latents_independently(self, latents, reference_flow, inject_noise=False):
110
  _, _, H, W = reference_flow.size()
111
+ b, _, f, h, w = latents.size()
112
  assert b == 1
113
  coords0 = coords_grid(f, H, W, device=latents.device).to(latents.dtype)
 
114
 
115
+ coords_t0 = coords0 + reference_flow
116
  coords_t0[:, 0] /= W
117
  coords_t0[:, 1] /= H
118
+
119
  coords_t0 = coords_t0 * 2.0 - 1.0
120
 
121
  coords_t0 = T.Resize((h, w))(coords_t0)
 
123
  coords_t0 = rearrange(coords_t0, 'f c h w -> f h w c')
124
 
125
  latents_0 = rearrange(latents[0], 'c f h w -> f c h w')
 
126
  warped = grid_sample(latents_0, coords_t0,
127
  mode='nearest', padding_mode='reflection')
128
+
129
+ if inject_noise:
130
+ idx = torch.logical_or(coords_t0 >= 1, coords_t0 < -1)
131
+ reset_noise = torch.randn(idx.shape)
132
+ idx = torch.logical_or(idx[:, :, :, 0], idx[:, :, :, 1])
133
+ idx = repeat(idx, "f w h -> f c w h", c=warped.shape[1])
134
+ reset_noise = torch.randn(
135
+ size=warped.shape, dtype=warped.dtype, device=warped.device)
136
+ warped[idx] = reset_noise[idx]
137
+
138
  warped = rearrange(warped, '(b f) c h w -> b c f h w', f=f)
139
  return warped
140
 
141
+ def DDIM_backward(self, num_inference_steps, timesteps, skip_t, t0, t1, do_classifier_free_guidance, null_embs, text_embeddings, latents_local,
142
+ latents_dtype, guidance_scale, guidance_stop_step, callback, callback_steps, extra_step_kwargs, num_warmup_steps):
143
  entered = False
144
+
145
  f = latents_local.shape[2]
146
+
147
+ latents_local = rearrange(latents_local, "b c f w h -> (b f) c w h")
148
+
149
  latents = latents_local.detach().clone()
150
  x_t0_1 = None
151
  x_t1_1 = None
 
 
152
 
153
  with self.progress_bar(total=num_inference_steps) as progress_bar:
154
  for i, t in enumerate(timesteps):
 
171
  with torch.no_grad():
172
  if null_embs is not None:
173
  text_embeddings[0] = null_embs[i][0]
174
+ te = torch.cat([repeat(text_embeddings[0, :, :], "c k -> f c k", f=f),
175
+ repeat(text_embeddings[1, :, :], "c k -> f c k", f=f)])
176
  noise_pred = self.unet(
177
  latent_model_input, t, encoder_hidden_states=te).sample.to(dtype=latents_dtype)
178
 
 
203
  if callback is not None and i % callback_steps == 0:
204
  callback(i, t, latents)
205
 
206
+ latents = rearrange(latents, "(b f) c w h -> b c f w h", f=f)
207
+
 
 
 
208
  res = {"x0": latents.detach().clone()}
209
  if x_t0_1 is not None:
210
+ x_t0_1 = rearrange(x_t0_1, "(b f) c w h -> b c f w h", f=f)
211
  res["x_t0_1"] = x_t0_1.detach().clone()
212
  if x_t1_1 is not None:
213
+ x_t1_1 = rearrange(x_t1_1, "(b f) c w h -> b c f w h", f=f)
214
  res["x_t1_1"] = x_t1_1.detach().clone()
215
  return res
216
 
 
222
  video = rearrange(video, "(b f) c h w -> b c f h w", f=video_length)
223
  video = (video / 2 + 0.5).clamp(0, 1)
224
  # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
225
+ video = video.detach().cpu()
226
  return video
227
 
228
+ def create_motion_field(self, motion_field_strength_x, motion_field_strength_y, frame_ids, video_length, latents):
229
+
230
+ reference_flow = torch.zeros(
231
+ (video_length-1, 2, 512, 512), device=latents.device, dtype=latents.dtype)
232
+ for fr_idx in range(video_length-1):
233
+ reference_flow[fr_idx, 0, :,
234
+ :] = motion_field_strength_x*(frame_ids[fr_idx]+1)
235
+ reference_flow[fr_idx, 1, :,
236
+ :] = motion_field_strength_y*(frame_ids[fr_idx]+1)
237
+ return reference_flow
238
 
239
+ def create_motion_field_and_warp_latents(self, motion_field_strength_x, motion_field_strength_y, frame_ids, video_length, inject_noise_to_warp, latents):
240
+
241
+ motion_field = self.create_motion_field(motion_field_strength_x=motion_field_strength_x,
242
+ motion_field_strength_y=motion_field_strength_y, latents=latents, video_length=video_length, frame_ids=frame_ids)
243
+ for idx, latent in enumerate(latents):
244
+ latents[idx] = self.warp_latents_independently(
245
+ latent[None], motion_field, inject_noise=inject_noise_to_warp)
246
+ return motion_field, latents
247
 
248
  @torch.no_grad()
249
  def __call__(
 
262
  List[torch.Generator]]] = None,
263
  xT: Optional[torch.FloatTensor] = None,
264
  null_embs: Optional[torch.FloatTensor] = None,
 
265
  motion_field_strength_x: float = 12,
266
+ motion_field_strength_y: float = 12,
267
  output_type: Optional[str] = "tensor",
268
  return_dict: bool = True,
269
  callback: Optional[Callable[[
270
  int, int, torch.FloatTensor], None]] = None,
271
  callback_steps: Optional[int] = 1,
272
  use_motion_field: bool = True,
273
+ smooth_bg: bool = False,
274
  smooth_bg_strength: float = 0.4,
275
+ inject_noise_to_warp: bool = False,
276
+ t0: int = 44,
277
+ t1: int = 47,
278
  **kwargs,
279
  ):
280
+ frame_ids = kwargs.pop("frame_ids", list(range(video_length)))
281
+
282
+ assert num_videos_per_prompt == 1
283
+ assert isinstance(prompt, list) and len(prompt) > 0
284
+ assert isinstance(negative_prompt, list) or negative_prompt is None
285
+
286
+ prompt_types = [prompt, negative_prompt]
287
+
288
+ for idx, prompt_type in enumerate(prompt_types):
289
+ prompt_template = None
290
+ for prompt in prompt_type:
291
+ if prompt_template is None:
292
+ prompt_template = prompt
293
+ else:
294
+ assert prompt == prompt_template
295
+ if prompt_types[idx] is not None:
296
+ prompt_types[idx] = prompt_types[idx][0]
297
+ prompt = prompt_types[0]
298
+ negative_prompt = prompt_types[1]
299
+
300
+ print(
301
+ f" Motion field strength x = {motion_field_strength_x}, y = {motion_field_strength_y}")
302
  print(f" Use: Motion field = {use_motion_field}")
303
  print(f" Use: Background smoothing = {smooth_bg}")
304
+ print(f"Inject noise to warp = {inject_noise_to_warp}")
305
  # Default height and width to unet
306
  height = height or self.unet.config.sample_size * self.vae_scale_factor
307
  width = width or self.unet.config.sample_size * self.vae_scale_factor
 
321
  text_embeddings = self._encode_prompt(
322
  prompt, device, num_videos_per_prompt, do_classifier_free_guidance, negative_prompt
323
  )
324
+
325
  # Prepare timesteps
326
  self.scheduler.set_timesteps(num_inference_steps, device=device)
327
  timesteps = self.scheduler.timesteps
328
+
329
  # print(f" Latent shape = {latents.shape}")
330
 
331
  # Prepare latent variables
 
334
  xT = self.prepare_latents(
335
  batch_size * num_videos_per_prompt,
336
  num_channels_latents,
337
+ 1,
338
  height,
339
  width,
340
  text_embeddings.dtype,
 
361
  None,
362
  )
363
  xT = torch.cat([xT, xT_missing], dim=2)
 
364
 
365
  xInit = xT.clone()
366
+
367
+ timesteps_ddpm = [981, 961, 941, 921, 901, 881, 861, 841, 821, 801, 781, 761, 741, 721,
368
+ 701, 681, 661, 641, 621, 601, 581, 561, 541, 521, 501, 481, 461, 441,
369
+ 421, 401, 381, 361, 341, 321, 301, 281, 261, 241, 221, 201, 181, 161,
370
+ 141, 121, 101, 81, 61, 41, 21, 1]
371
+ timesteps_ddpm.reverse()
372
+
373
+ t0 = timesteps_ddpm[t0]
374
+ t1 = timesteps_ddpm[t1]
375
+ print(f"t0 = {t0} t1 = {t1}")
376
  x_t1_1 = None
377
 
 
378
  # Prepare extra step kwargs.
379
  extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
380
  # Denoising loop
381
  num_warmup_steps = len(timesteps) - \
382
  num_inference_steps * self.scheduler.order
 
383
 
384
+ shape = (batch_size, num_channels_latents, 1, height //
385
+ self.vae_scale_factor, width // self.vae_scale_factor)
386
+ if inject_noise_to_warp and use_motion_field:
387
+ # if we inject to noise to warp function, we do it for timesteps T = 1000
388
+
389
+ x_t0_k = xT[:, :, :1, :, :].repeat(1, 1, video_length-1, 1, 1)
390
+
391
+ # reference_flow, x_t0_k = self.create_motion_field_and_warp_latents(motion_field_strength_x=motion_field_strength_x, motion_field_strength_y=motion_field_strength_y,
392
+ # frame_ids=frame_ids,video_length=video_length,inject_noise_to_warp=inject_noise_to_warp,latents = x_t0_k)
393
+ # xT =torch.cat([xT, x_t0_k], dim=2).clone().detach()
394
 
395
  ddim_res = self.DDIM_backward(num_inference_steps=num_inference_steps, timesteps=timesteps, skip_t=1000, t0=t0, t1=t1, do_classifier_free_guidance=do_classifier_free_guidance,
396
+ null_embs=null_embs, text_embeddings=text_embeddings, latents_local=xT, latents_dtype=dtype, guidance_scale=guidance_scale, guidance_stop_step=guidance_stop_step,
397
+ callback=callback, callback_steps=callback_steps, extra_step_kwargs=extra_step_kwargs, num_warmup_steps=num_warmup_steps)
398
+
399
  x0 = ddim_res["x0"].detach()
400
+
401
  if "x_t0_1" in ddim_res:
402
  x_t0_1 = ddim_res["x_t0_1"].detach()
403
  if "x_t1_1" in ddim_res:
 
405
  del ddim_res
406
  del xT
407
 
408
+ if inject_noise_to_warp and use_motion_field:
409
+ # DDPM forward to allow for more motion
410
+ if t1 > t0:
411
+ x_t1_k = self.DDPM_forward(
412
+ x0=x_t0_1, t0=t0, tMax=t1, device=device, shape=shape, text_embeddings=text_embeddings, generator=generator)
413
+ else:
414
+ x_t1_k = x_t0_k
415
+
416
+ if x_t1_1 is None:
417
+ raise Exception
418
+
419
+ x_t1 = x_t1_k.clone().detach()
420
+
421
+ ddim_res = self.DDIM_backward(num_inference_steps=num_inference_steps, timesteps=timesteps, skip_t=t1, t0=-1, t1=-1, do_classifier_free_guidance=do_classifier_free_guidance,
422
+ null_embs=null_embs, text_embeddings=text_embeddings, latents_local=x_t1, latents_dtype=dtype, guidance_scale=guidance_scale, guidance_stop_step=guidance_stop_step,
423
+ callback=callback, callback_steps=callback_steps, extra_step_kwargs=extra_step_kwargs, num_warmup_steps=num_warmup_steps)
424
+
425
+ x0 = ddim_res["x0"].detach()
426
+ del ddim_res
427
+ del x_t1
428
+ del x_t1_k
429
+
430
+ if use_motion_field and not inject_noise_to_warp:
431
  del x0
 
 
 
 
 
432
 
433
+ x_t0_k = x_t0_1[:, :, :1, :, :].repeat(1, 1, video_length-1, 1, 1)
 
 
 
 
 
 
434
 
435
+ reference_flow, x_t0_k = self.create_motion_field_and_warp_latents(
436
+ motion_field_strength_x=motion_field_strength_x, motion_field_strength_y=motion_field_strength_y, latents=x_t0_k, video_length=video_length,
437
+ inject_noise_to_warp=inject_noise_to_warp, frame_ids=frame_ids)
438
 
439
  # assuming t0=t1=1000, if t0 = 1000
440
  if t1 > t0:
 
449
  x_t1 = torch.cat([x_t1_1, x_t1_k], dim=2).clone().detach()
450
 
451
  ddim_res = self.DDIM_backward(num_inference_steps=num_inference_steps, timesteps=timesteps, skip_t=t1, t0=-1, t1=-1, do_classifier_free_guidance=do_classifier_free_guidance,
452
+ null_embs=null_embs, text_embeddings=text_embeddings, latents_local=x_t1, latents_dtype=dtype, guidance_scale=guidance_scale,
453
+ guidance_stop_step=guidance_stop_step, callback=callback, callback_steps=callback_steps, extra_step_kwargs=extra_step_kwargs, num_warmup_steps=num_warmup_steps)
454
 
455
  x0 = ddim_res["x0"].detach()
456
  del ddim_res
457
+ del x_t1
458
+ del x_t1_1
459
+ del x_t1_k
460
+
461
  else:
462
  x_t1 = x_t1_1.clone()
463
+ x_t1_1 = x_t1_1[:, :, :1, :, :].clone()
464
+ x_t1_k = x_t1_1[:, :, 1:, :, :].clone()
465
  x_t0_k = x_t0_1[:, :, 1:, :, :].clone()
466
+ x_t0_1 = x_t0_1[:, :, :1, :, :].clone()
467
 
468
  # smooth background
469
  if smooth_bg:
 
485
  mask = dilation(mask[None].to(x0.device), kernel)[0]
486
  M_FG[batch_idx, frame_idx, :, :] = mask
487
 
 
488
  x_t1_1_fg_masked = x_t1_1 * \
489
  (1 - repeat(M_FG[:, 0, :, :],
490
  "b w h -> b c 1 w h", c=x_t1_1.shape[1]))
491
 
 
492
  x_t1_1_fg_masked_moved = []
493
  for batch_idx, x_t1_1_fg_masked_b in enumerate(x_t1_1_fg_masked):
494
  x_t1_fg_masked_b = x_t1_1_fg_masked_b.clone()
 
498
  if use_motion_field:
499
  x_t1_fg_masked_b = x_t1_fg_masked_b[None]
500
  x_t1_fg_masked_b = self.warp_latents_independently(
501
+ x_t1_fg_masked_b, reference_flow, inject_noise=False)
502
  else:
503
  x_t1_fg_masked_b = x_t1_fg_masked_b[None]
504
 
 
514
  for batch_idx, m_fg_1_b in enumerate(M_FG_1):
515
  m_fg_1_b = m_fg_1_b[None, None]
516
  m_fg_b = m_fg_1_b.repeat(1, 1, video_length-1, 1, 1)
517
+ if use_motion_field:
518
  m_fg_b = self.warp_latents_independently(
519
+ m_fg_b.clone(), reference_flow, inject_noise=False)
520
  M_FG_warped.append(
521
  torch.cat([m_fg_1_b[:1, 0], m_fg_b[:1, 0]], dim=1))
522
 
 
527
  M_BG = (1-M_FG) * (1 - M_FG_warped)
528
  M_BG = repeat(M_BG, "b f h w -> b c f h w", c=channels)
529
  a_convex = smooth_bg_strength
 
 
 
530
 
531
+ latents = (1-M_BG) * x_t1 + M_BG * (a_convex *
532
+ x_t1 + (1-a_convex) * x_t1_1_fg_masked_moved)
 
 
 
 
533
 
534
  ddim_res = self.DDIM_backward(num_inference_steps=num_inference_steps, timesteps=timesteps, skip_t=t1, t0=-1, t1=-1, do_classifier_free_guidance=do_classifier_free_guidance,
535
+ null_embs=null_embs, text_embeddings=text_embeddings, latents_local=latents, latents_dtype=dtype, guidance_scale=guidance_scale,
536
+ guidance_stop_step=guidance_stop_step, callback=callback, callback_steps=callback_steps, extra_step_kwargs=extra_step_kwargs, num_warmup_steps=num_warmup_steps)
537
  x0 = ddim_res["x0"].detach()
538
  del ddim_res
539
+ del latents
540
+
541
+ latents = x0
542
+
543
+ # manually for max memory savings
544
+ if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
545
+ self.unet.to("cpu")
546
+ torch.cuda.empty_cache()
547
+
548
+ if output_type == "latent":
549
+ image = latents
550
+ has_nsfw_concept = None
551
+ else:
552
+ image = self.decode_latents(latents)
553
+
554
+ # Run safety checker
555
+ image, has_nsfw_concept = self.run_safety_checker(
556
+ image, device, text_embeddings.dtype)
557
+ image = rearrange(image, "b c f h w -> (b f) h w c")
558
 
559
+ # Offload last model to CPU
560
+ if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
561
+ self.final_offload_hook.offload()
562
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
563
  if not return_dict:
564
+ return (image, has_nsfw_concept)
565
 
566
+ return TextToVideoPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
utils.py CHANGED
@@ -1,8 +1,10 @@
1
  import os
 
 
2
  import numpy as np
3
  import torch
4
  import torchvision
5
- from torchvision.transforms import Resize
6
  import imageio
7
  from einops import rearrange
8
  import cv2
@@ -11,23 +13,33 @@ from annotator.util import resize_image, HWC3
11
  from annotator.canny import CannyDetector
12
  from annotator.openpose import OpenposeDetector
13
  import decord
14
- decord.bridge.set_bridge('torch')
15
 
16
  apply_canny = CannyDetector()
17
  apply_openpose = OpenposeDetector()
18
 
19
 
20
- def add_watermark(image, im_size_h, im_size_w, watermark_path="__assets__/picsart_watermark.jpg",
21
- wmsize=16, bbuf=5, opacity=0.9):
22
  '''
23
  Creates a watermark on the saved inference image.
24
  We request that you do not remove this to properly assign credit to
25
  Shi-Lab's work.
26
  '''
27
- watermark = Image.open(watermark_path).resize((wmsize, wmsize))
28
- loc_h = im_size_h - wmsize - bbuf
29
- loc_w = im_size_w - wmsize - bbuf
30
- image[loc_h:-bbuf, loc_w:-bbuf, :] = watermark
 
 
 
 
 
 
 
 
 
 
 
31
  return image
32
 
33
 
@@ -61,7 +73,7 @@ def pre_process_pose(input_video, apply_pose_detect: bool = True):
61
  return rearrange(control, 'f h w c -> f c h w')
62
 
63
 
64
- def create_video(frames, fps, rescale=False, path=None):
65
  if path is None:
66
  dir = "temporal"
67
  os.makedirs(dir, exist_ok=True)
@@ -74,18 +86,19 @@ def create_video(frames, fps, rescale=False, path=None):
74
  x = (x + 1.0) / 2.0 # -1,1 -> 0,1
75
  x = (x * 255).numpy().astype(np.uint8)
76
 
77
- h_, w_, _ = x.shape
78
- x = add_watermark(x, im_size_h=h_, im_size_w=w_)
79
  outputs.append(x)
80
  # imageio.imsave(os.path.join(dir, os.path.splitext(name)[0] + f'_{i}.jpg'), x)
81
 
82
  imageio.mimsave(path, outputs, fps=fps)
83
  return path
84
 
85
- def create_gif(frames, fps, rescale=False):
86
- dir = "temporal"
87
- os.makedirs(dir, exist_ok=True)
88
- path = os.path.join(dir, 'canny_db.gif')
 
89
 
90
  outputs = []
91
  for i, x in enumerate(frames):
@@ -93,8 +106,8 @@ def create_gif(frames, fps, rescale=False):
93
  if rescale:
94
  x = (x + 1.0) / 2.0 # -1,1 -> 0,1
95
  x = (x * 255).numpy().astype(np.uint8)
96
- h_, w_, _ = x.shape
97
- x = add_watermark(x, im_size_h=h_, im_size_w=w_)
98
  outputs.append(x)
99
  # imageio.imsave(os.path.join(dir, os.path.splitext(name)[0] + f'_{i}.jpg'), x)
100
 
@@ -103,7 +116,6 @@ def create_gif(frames, fps, rescale=False):
103
 
104
  def prepare_video(video_path:str, resolution:int, device, dtype, normalize=True, start_t:float=0, end_t:float=-1, output_fps:int=-1):
105
  vr = decord.VideoReader(video_path)
106
- video = vr.get_batch(range(0, len(vr))).asnumpy()
107
  initial_fps = vr.get_avg_fps()
108
  if output_fps == -1:
109
  output_fps = int(initial_fps)
@@ -113,24 +125,27 @@ def prepare_video(video_path:str, resolution:int, device, dtype, normalize=True,
113
  end_t = min(len(vr) / initial_fps, end_t)
114
  assert 0 <= start_t < end_t
115
  assert output_fps > 0
116
- f, h, w, c = video.shape
117
  start_f_ind = int(start_t * initial_fps)
118
  end_f_ind = int(end_t * initial_fps)
119
  num_f = int((end_t - start_t) * output_fps)
120
  sample_idx = np.linspace(start_f_ind, end_f_ind, num_f, endpoint=False).astype(int)
121
- video = video[sample_idx]
 
 
 
 
 
122
  video = rearrange(video, "f h w c -> f c h w")
123
  video = torch.Tensor(video).to(device).to(dtype)
124
  if h > w:
125
  w = int(w * resolution / h)
126
  w = w - w % 8
127
  h = resolution - resolution % 8
128
- video = Resize((h, w))(video)
129
  else:
130
  h = int(h * resolution / w)
131
  h = h - h % 8
132
  w = resolution - resolution % 8
133
- video = Resize((h, w))(video)
134
  if normalize:
135
  video = video / 127.5 - 1.0
136
  return video, output_fps
 
1
  import os
2
+
3
+ import PIL.Image
4
  import numpy as np
5
  import torch
6
  import torchvision
7
+ from torchvision.transforms import Resize, InterpolationMode
8
  import imageio
9
  from einops import rearrange
10
  import cv2
 
13
  from annotator.canny import CannyDetector
14
  from annotator.openpose import OpenposeDetector
15
  import decord
16
+ # decord.bridge.set_bridge('torch')
17
 
18
  apply_canny = CannyDetector()
19
  apply_openpose = OpenposeDetector()
20
 
21
 
22
+ def add_watermark(image, watermark_path, wm_rel_size=1/16, boundary=5):
 
23
  '''
24
  Creates a watermark on the saved inference image.
25
  We request that you do not remove this to properly assign credit to
26
  Shi-Lab's work.
27
  '''
28
+ watermark = Image.open(watermark_path)
29
+ w_0, h_0 = watermark.size
30
+ H, W, _ = image.shape
31
+ wmsize = int(max(H, W) * wm_rel_size)
32
+ aspect = h_0 / w_0
33
+ if aspect > 1.0:
34
+ watermark = watermark.resize((wmsize, int(aspect * wmsize)), Image.LANCZOS)
35
+ else:
36
+ watermark = watermark.resize((int(wmsize / aspect), wmsize), Image.LANCZOS)
37
+ w, h = watermark.size
38
+ loc_h = H - h - boundary
39
+ loc_w = W - w - boundary
40
+ image = Image.fromarray(image)
41
+ mask = watermark if watermark.mode in ('RGBA', 'LA') else None
42
+ image.paste(watermark, (loc_w, loc_h), mask)
43
  return image
44
 
45
 
 
73
  return rearrange(control, 'f h w c -> f c h w')
74
 
75
 
76
+ def create_video(frames, fps, rescale=False, path=None, watermark=None):
77
  if path is None:
78
  dir = "temporal"
79
  os.makedirs(dir, exist_ok=True)
 
86
  x = (x + 1.0) / 2.0 # -1,1 -> 0,1
87
  x = (x * 255).numpy().astype(np.uint8)
88
 
89
+ if watermark is not None:
90
+ x = add_watermark(x, watermark)
91
  outputs.append(x)
92
  # imageio.imsave(os.path.join(dir, os.path.splitext(name)[0] + f'_{i}.jpg'), x)
93
 
94
  imageio.mimsave(path, outputs, fps=fps)
95
  return path
96
 
97
+ def create_gif(frames, fps, rescale=False, path=None, watermark=None):
98
+ if path is None:
99
+ dir = "temporal"
100
+ os.makedirs(dir, exist_ok=True)
101
+ path = os.path.join(dir, 'canny_db.gif')
102
 
103
  outputs = []
104
  for i, x in enumerate(frames):
 
106
  if rescale:
107
  x = (x + 1.0) / 2.0 # -1,1 -> 0,1
108
  x = (x * 255).numpy().astype(np.uint8)
109
+ if watermark is not None:
110
+ x = add_watermark(x, watermark)
111
  outputs.append(x)
112
  # imageio.imsave(os.path.join(dir, os.path.splitext(name)[0] + f'_{i}.jpg'), x)
113
 
 
116
 
117
  def prepare_video(video_path:str, resolution:int, device, dtype, normalize=True, start_t:float=0, end_t:float=-1, output_fps:int=-1):
118
  vr = decord.VideoReader(video_path)
 
119
  initial_fps = vr.get_avg_fps()
120
  if output_fps == -1:
121
  output_fps = int(initial_fps)
 
125
  end_t = min(len(vr) / initial_fps, end_t)
126
  assert 0 <= start_t < end_t
127
  assert output_fps > 0
 
128
  start_f_ind = int(start_t * initial_fps)
129
  end_f_ind = int(end_t * initial_fps)
130
  num_f = int((end_t - start_t) * output_fps)
131
  sample_idx = np.linspace(start_f_ind, end_f_ind, num_f, endpoint=False).astype(int)
132
+ video = vr.get_batch(sample_idx)
133
+ if torch.is_tensor(video):
134
+ video = video.detach().cpu().numpy()
135
+ else:
136
+ video = video.asnumpy()
137
+ _, h, w, _ = video.shape
138
  video = rearrange(video, "f h w c -> f c h w")
139
  video = torch.Tensor(video).to(device).to(dtype)
140
  if h > w:
141
  w = int(w * resolution / h)
142
  w = w - w % 8
143
  h = resolution - resolution % 8
 
144
  else:
145
  h = int(h * resolution / w)
146
  h = h - h % 8
147
  w = resolution - resolution % 8
148
+ video = Resize((h, w), interpolation=InterpolationMode.BILINEAR, antialias=True)(video)
149
  if normalize:
150
  video = video / 127.5 - 1.0
151
  return video, output_fps