File size: 1,602 Bytes
d6f0309
 
ab9d20f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
V0_1: 2,800 more training steps compared to V0.

Current V0 went through : 
- 100 initial steps at LR 5e-5 (batch size 8)
- 1100 more steps at LR 1e-4 (batch size 32)
- 600 more steps at LR 1e-4 (batch size 8)
- 600 more steps at LR 7.5e-5 (batch size 8)

Dataset has 30,000 examples covering:

- Text-to-video with no reference: 6,500
- Text-to-image with no reference: 3,500

- Still-image-to-video using:
  - First frame from the target video: 750
  - Middle frame from the target video: 750
  - Last frame from the target video: 750
  - Frame from the same source but outside the target interval: 750

- Face-reference-to-video using a close-up crop of the primary face: 3,000

- Image-to-image using a different image from the same gallery: 3,000

- Image restoration and inpainting:
  - Masked image reference: 800
  - Degraded image reference: 800

- Image outpainting using an aggressively cropped reference: 2,400

- Video outpainting:
  - Static spatial crop: 501
  - Tracked spatial crop: 499

- Temporal video completion:
  - Reference video immediately before the target: 467
  - Reference segment inside the complete target: 467
  - Reference video immediately after the target: 466

- Audio-to-video using the target audio as conditioning: 2,000

- Video-to-audio / sound generation using a muted,
  low-resolution video reference: 1,400

- Structural conditioning:
  - Single-frame depth: 160
  - Single-frame optical flow: 147
  - Single-frame pose: 145
  - Single-frame segmentation: 148
  - Video depth: 146
  - Video optical flow: 153
  - Video pose: 149
  - Video segmentation: 152