VIDEOCOMPOSER: COMPOSITIONAL VIDEO SYNTHESIS WITH MOTION CONTROLLABILITY��VIDEOCOMPOSER:具有運動可控性的影片合成
Presenter: Hsin-Wei, Lin
Date: 2024/01/10
OUTLINE
INTRODUCTION (1/2)
3
現有的DM影片生成調整方法
Composer
VideoComposer
INTRODUCTION (2/2) – 貢獻
4
RELATED WORK
5
GAN
VAE
flow models
Diffusion
Enhanced�&�Improves
text
encoder
video
generation
DALL· E [CFG + CLIP]
Imagen [LLM: T5]
pix2pix-zero, prompt-to-prompt editing
[交叉注意力以遵循指導文本]
ControlNet, T2I-adapter
[合併空間條件至模型中]
multi-condition
Composer
Composer
image-to-vid
others
VideoComposer
WORK FLOW (1/7) – 系統架構圖
6
3
2
1
WORK FLOW (2/7) – VLDM (1/2)
7
Video
Pre-trained�Encoder
Latent�Representation
Decoder
D
Pixel Space
WORK FLOW (3/7) – VLDM (2/2)
8
T = 1000
Latent
“3D U-Net”
Conditions
WORK FLOW (4/7) – 條件 (1/2)
9
圖源:https://github.com/mlfoundations/open_clip
WORK FLOW (5/7) – 條件 (2/2)
10
圖源:Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-shot Cross-dataset Transfer
WORK FLOW (6/7) – STC-ENCODER
11
WORK FLOW (7/7) – 訓練和推理
12
EXPERIMENTS (1/12)– 實驗設定
13
EXPERIMENTS (2/12) – 具有多條件的可組合影片生成(1/6)
14
“Tiger”
“Tiger”
TEXT + SINGLE IMAGE
TEXT + SINGLE IMAGE + MOTION
EXPERIMENTS (3/12) – 具有多條件的可組合影片生成(2/6)
15
“Bouquet of the yellow leaves.
Autumn girl walking in city park”
TEXT + SINGLE IMAGE
TEXT + SINGLE IMAGE + DEPTH
“Bouquet of the yellow leaves.
Autumn girl walking in city park”
EXPERIMENTS (4/12) – 具有多條件的可組合影片生成(3/6)
16
“The Matterhorn with flowers in the foreground”
TEXT + SINGLE IMAGE
TEXT + SINGLE IMAGE + SKETCH
“The Matterhorn with flowers in the foreground”
EXPERIMENTS (5/12) – 具有多條件的可組合影片生成(4/6)
17
INPUT MASKED VIDEO
TEXT + MASK
TEXT + MASK + DEPTH
TEXT + MASK + SKETCH
“Swan lazily drifting and trimming itself on a calm river”
“Swan lazily drifting and trimming itself on a calm river”
“Black swan lazily drifting and trimming itself on a calm river”
選取欲修改區域之遮罩
EXPERIMENTS (6/12) – 具有多條件的可組合影片生成(5/6)
18
INPUT MASKED VIDEO
TEXT + MASK
TEXT + MASK + DEPTH
TEXT + MASK + SKETCH
“Coastal view of Cleveland Road and Robin Hood's Bay, blue sky and sea with white clouds”
“Coastal view of Cleveland Road and Robin Hood's Bay, blue sky and sea with white clouds”
EXPERIMENTS (7/12) – 具有多條件的可組合影片生成(6/6)
19
INPUT HAND DRAWINGS
TEXT + SINGLE SKETCH
TEXT + SINGLE SKETCH + DEPTH
“Pigeon sits on a stone with an iron net behind it”
“Pigeon sits on a stone with an iron net behind it”
單個草圖的合成訓練使VideoComposer具有動畫化靜態草圖的能力
EXPERIMENTS (8/12) – 運動控制的實驗結果(1/3)
20
EXPERIMENTS (9/12) – 運動控制的實驗結果(2/3)
21
“American Staffordshire Terrier is lying on the floor in an abandoned building.”
EXPERIMENTS (10/12) – 運動控制的實驗結果(3/3)
22
“A moving golden moon”
TEXT + HAND-CRAFTED MOTIONS
TEXT + HAND-CRAFTED MOTIONS
TEXT + HAND-CRAFTED MOTIONS
TEXT + HAND-CRAFTED MOTIONS
EXPERIMENTS (11/12) – 消融研究(1/2)
23
EXPERIMENTS (12/12) – 消融研究(2/2)
24
CONCLUSION
25
THAT’S ALL�THANK YOU!