clark-labs commited on
Commit
362ddaa
·
verified ·
1 Parent(s): 8476529

Add true 200px compressed aligned video demo

Browse files
.gitattributes CHANGED
@@ -56,3 +56,12 @@ examples/seedance-action/seedance-2-action-bicubic-x2-400p.mp4 filter=lfs diff=l
56
  examples/seedance-action/seedance-2-action-clark-x2-400p.mp4 filter=lfs diff=lfs merge=lfs -text
57
  examples/seedance-action/seedance-2-action-held-out-target-400p.mp4 filter=lfs diff=lfs merge=lfs -text
58
  examples/seedance-action/seedance-2-action-real-input-200p-crf32.mp4 filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
 
 
 
 
 
56
  examples/seedance-action/seedance-2-action-clark-x2-400p.mp4 filter=lfs diff=lfs merge=lfs -text
57
  examples/seedance-action/seedance-2-action-held-out-target-400p.mp4 filter=lfs diff=lfs merge=lfs -text
58
  examples/seedance-action/seedance-2-action-real-input-200p-crf32.mp4 filter=lfs diff=lfs merge=lfs -text
59
+ examples/seedance-action/seedance-2-action-held-out-target-400px.mp4 filter=lfs diff=lfs merge=lfs -text
60
+ examples/seedance-action/seedance-2-action-real-input-200px-crf36.mp4 filter=lfs diff=lfs merge=lfs -text
61
+ examples/seedance-action/seedance-2-action-bicubic-x2-400px.mp4 filter=lfs diff=lfs merge=lfs -text
62
+ examples/seedance-action/seedance-2-action-clark-x2-400px.mp4 filter=lfs diff=lfs merge=lfs -text
63
+ examples/seedance-action/seedance-2-action-200px-triptych.mp4 filter=lfs diff=lfs merge=lfs -text
64
+ examples/seedance-action/seedance-2-action-200px-before-after.mp4 filter=lfs diff=lfs merge=lfs -text
65
+ examples/seedance-action/seedance-2-action-200px-before-after-poster.png filter=lfs diff=lfs merge=lfs -text
66
+ examples/seedance-action/seedance-2-action-200px-wipe.mp4 filter=lfs diff=lfs merge=lfs -text
67
+ examples/seedance-action/seedance-2-action-200px-wipe-poster.png filter=lfs diff=lfs merge=lfs -text
README.md CHANGED
@@ -21,37 +21,40 @@ The browser artifact has 8,496 parameters and runs as three fused FP16 WebGPU re
21
 
22
  ### Moving Seedance 2.0 action test
23
 
24
- <video controls loop playsinline muted width="100%" poster="examples/seedance-action/seedance-2-action-200p-before-after-poster.png">
25
- <source src="https://huggingface.co/clark-labs/clark-realtime-ui-video-upscaler-x2/resolve/main/examples/seedance-action/seedance-2-action-200p-before-after-magnified.mp4" type="video/mp4">
26
  </video>
27
 
28
- [Watch the large 200p before/after](examples/seedance-action/seedance-2-action-200p-before-after-magnified.mp4) · [Inspect the 200p three-way control](examples/seedance-action/seedance-2-action-200p-magnified-detail.mp4) · [Watch the original Seedance master](examples/seedance-action/seedance-2-action-original-720p.mp4)
29
 
30
  This five-second, 24 FPS test starts with a public 1280×720 Seedance 2.0 action
31
  generation: a running puppy, tracking camera, fur, grass, flowers, and a moving
32
  butterfly. A deterministic product-demo HUD is added before downsampling so the
33
  test also contains the small labels, cards, and buttons this model targets. The
34
- large demo intentionally uses a heavily compressed 356×200 input and displays a
35
- moving detail crop at 2×. Clark's honest x2 output is 712×400; the demo does not
36
- pretend that x2 reaches the original 720p reference.
 
37
 
38
- Against a separately derived 712×400 held-out target, across all 120 frames:
39
 
40
  | Measure | Bicubic x2 | Clark x2 | Gain |
41
  |---|---:|---:|---:|
42
- | RGB PSNR | 27.054 dB | 27.169 dB | +0.115 dB |
43
- | Edge PSNR | 14.915 dB | 16.137 dB | +1.222 dB |
44
- | Temporal-delta PSNR | 27.891 dB | 27.744 dB | -0.147 dB |
45
-
46
- The two-panel hero makes the low-resolution input-to-output change legible. The
47
- three-way control remains the scientific comparison: Clark improves edge
48
- recovery over bicubic, but only narrowly improves whole-frame reconstruction
49
- and still trails bicubic on temporal-delta fidelity. The source, separate
 
50
  bicubic and Clark outputs, held-out reference, exact prompt, seed, hashes,
51
- crop, and metric definitions are all included under
52
  `examples/seedance-action/` and `technical/seedance-action-video-evaluation.json`.
53
 
54
- The less severe 640×360 to 1280×720 test is also included as
 
55
  [a native triptych](examples/seedance-action/seedance-2-action-native-triptych.mp4).
56
  It measures +0.703 dB RGB PSNR, +2.673 dB edge PSNR, and -0.922 dB
57
  temporal-delta PSNR versus bicubic.
 
21
 
22
  ### Moving Seedance 2.0 action test
23
 
24
+ <video controls loop playsinline muted width="100%" poster="examples/seedance-action/seedance-2-action-200px-wipe-poster.png">
25
+ <source src="https://huggingface.co/clark-labs/clark-realtime-ui-video-upscaler-x2/resolve/main/examples/seedance-action/seedance-2-action-200px-wipe.mp4" type="video/mp4">
26
  </video>
27
 
28
+ [Watch the aligned 200-pixel wipe](examples/seedance-action/seedance-2-action-200px-wipe.mp4) · [Inspect the source/bicubic/Clark control](examples/seedance-action/seedance-2-action-200px-triptych.mp4) · [Download the real 200×112 input](examples/seedance-action/seedance-2-action-real-input-200px-crf36.mp4) · [Watch the original Seedance master](examples/seedance-action/seedance-2-action-original-720p.mp4)
29
 
30
  This five-second, 24 FPS test starts with a public 1280×720 Seedance 2.0 action
31
  generation: a running puppy, tracking camera, fur, grass, flowers, and a moving
32
  butterfly. A deterministic product-demo HUD is added before downsampling so the
33
  test also contains the small labels, cards, and buttons this model targets. The
34
+ hero intentionally uses a heavily compressed, genuinely tiny 200×112 input.
35
+ The aligned wipe compares the same moving pixels: the decoded source is
36
+ nearest-neighbor magnified for inspection and Clark's honest 400×224 x2
37
+ output is magnified 2×. Display magnification is not counted as model output.
38
 
39
+ Against a separately derived 400×224 held-out target, across all 120 frames:
40
 
41
  | Measure | Bicubic x2 | Clark x2 | Gain |
42
  |---|---:|---:|---:|
43
+ | RGB PSNR | 23.383 dB | 23.338 dB | -0.045 dB |
44
+ | Edge PSNR | 12.296 dB | 12.538 dB | +0.241 dB |
45
+ | Temporal-delta PSNR | 24.784 dB | 24.666 dB | -0.119 dB |
46
+
47
+ The moving wipe makes the low-resolution input-to-output change legible because
48
+ the comparison is spatially aligned. The three-way control remains the
49
+ scientific comparison: at this deliberately extreme size and compression,
50
+ Clark improves edge recovery over bicubic but does not beat bicubic on
51
+ whole-frame or temporal-delta fidelity. The source, separate
52
  bicubic and Clark outputs, held-out reference, exact prompt, seed, hashes,
53
+ and metric definitions are all included under
54
  `examples/seedance-action/` and `technical/seedance-action-video-evaluation.json`.
55
 
56
+ The less severe 356×200 to 712×400 test improves RGB PSNR by 0.115 dB and edge
57
+ PSNR by 1.222 dB versus bicubic. The 640×360 to 1280×720 test is included as
58
  [a native triptych](examples/seedance-action/seedance-2-action-native-triptych.mp4).
59
  It measures +0.703 dB RGB PSNR, +2.673 dB edge PSNR, and -0.922 dB
60
  temporal-delta PSNR versus bicubic.
examples/seedance-action/seedance-2-action-200px-before-after-poster.png ADDED

Git LFS Details

  • SHA256: 57fef502d505a204b2ad06feba3a99c899417850650709c3844342bea739c165
  • Pointer size: 131 Bytes
  • Size of remote file: 412 kB
examples/seedance-action/seedance-2-action-200px-before-after.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:60f66dc27de71c8ea08866043df3fbc6231cb6e1606587bc5ae5233fdce0ee9d
3
+ size 3661573
examples/seedance-action/seedance-2-action-200px-triptych.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2d999796f28d5367789a65fb0393ac58d7f6d0ebe0a7c35197fcc40b4d8804c7
3
+ size 5677963
examples/seedance-action/seedance-2-action-200px-wipe-poster.png ADDED

Git LFS Details

  • SHA256: 91bfb43a0de46fcd5022025ef49f98e90f366332228b7aab4a744e84dd23563f
  • Pointer size: 131 Bytes
  • Size of remote file: 234 kB
examples/seedance-action/seedance-2-action-200px-wipe.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b24fc83fca9ac2c616c89bcd48f85f98bee819e2fc00917b7206855e45c2b3d6
3
+ size 1978956
examples/seedance-action/seedance-2-action-bicubic-x2-400px.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e09c6ff8d6535ccd0e8e1304e503bc3907251bdd9a71af5bb6139d12bd8ab868
3
+ size 1091381
examples/seedance-action/seedance-2-action-clark-x2-400px.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fc1f637292f725a7416479be08ae6eb879ac4584bf8be3fa9d949d521aa8ba68
3
+ size 1280429
examples/seedance-action/seedance-2-action-held-out-target-400px.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:065becf9e62e0456beb7d6474aa305761919e0b966abdc9cda3ddbb8302928e9
3
+ size 1401259
examples/seedance-action/seedance-2-action-real-input-200px-crf36.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:503e06db4b6e92a0f658d5acb0ad334c5f251d4d1e5bf22fe1993cfd0822e7ab
3
+ size 127966
technical/package-ledger.json CHANGED
@@ -3,8 +3,8 @@
3
  "checkpoint_sha256": "3cbb8a3358bbeaa239c517f1171dbb7706c2d453c961d17a4e6166aefde8c692",
4
  "artifacts": {
5
  "README.md": {
6
- "bytes": 3867,
7
- "sha256": "8b3831f2d9151b8f1bb97850b82ba061bb9a3c264f353ed1e4f88ac109446c7e"
8
  },
9
  "LICENSE": {
10
  "bytes": 11348,
 
3
  "checkpoint_sha256": "3cbb8a3358bbeaa239c517f1171dbb7706c2d453c961d17a4e6166aefde8c692",
4
  "artifacts": {
5
  "README.md": {
6
+ "bytes": 6497,
7
+ "sha256": "4e0afec1d8b6b96ae9b28aff63f13a06e840b8ed3a0b2ecb3e603104c2c378ce"
8
  },
9
  "LICENSE": {
10
  "bytes": 11348,
technical/seedance-action-video-evaluation.json CHANGED
@@ -59,6 +59,27 @@
59
  "temporal_gain_db": -0.14667097414172403
60
  }
61
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
62
  "artifacts": {
63
  "examples/seedance-action/seedance-2-action-original-720p.mp4": {
64
  "bytes": 1950352,
@@ -135,6 +156,42 @@
135
  "examples/seedance-action/seedance-2-action-200p-before-after-poster.png": {
136
  "bytes": 604436,
137
  "sha256": "623a7a89cfd43eca6ec0d5573236d888f58287e05eafe189abd492980483bbc0"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
138
  }
139
  }
140
  }
 
59
  "temporal_gain_db": -0.14667097414172403
60
  }
61
  },
62
+ "compressed_200px_wide_stress_test": {
63
+ "contract": {
64
+ "held_out_target": "400x224 Lanczos derivative of the 720p reference",
65
+ "real_input": "200x112 H.264 CRF 36 derivative of the held-out target",
66
+ "bicubic": "bicubic x2 from the decoded compressed input",
67
+ "clark": "RGB student x2 from the same decoded compressed input",
68
+ "display": "source nearest-neighbor 4x; x2 outputs nearest-neighbor 2x"
69
+ },
70
+ "metrics": {
71
+ "frames": 120,
72
+ "bicubic_psnr_db": 23.383082163063655,
73
+ "clark_psnr_db": 23.33781800943331,
74
+ "bicubic_edge_psnr_db": 12.296378722776026,
75
+ "clark_edge_psnr_db": 12.537776923891721,
76
+ "bicubic_temporal_psnr_db": 24.784364514745608,
77
+ "clark_temporal_psnr_db": 24.665802563480675,
78
+ "psnr_gain_db": -0.045264153630345305,
79
+ "edge_gain_db": 0.241398201115695,
80
+ "temporal_gain_db": -0.11856195126493319
81
+ }
82
+ },
83
  "artifacts": {
84
  "examples/seedance-action/seedance-2-action-original-720p.mp4": {
85
  "bytes": 1950352,
 
156
  "examples/seedance-action/seedance-2-action-200p-before-after-poster.png": {
157
  "bytes": 604436,
158
  "sha256": "623a7a89cfd43eca6ec0d5573236d888f58287e05eafe189abd492980483bbc0"
159
+ },
160
+ "examples/seedance-action/seedance-2-action-held-out-target-400px.mp4": {
161
+ "bytes": 1401259,
162
+ "sha256": "065becf9e62e0456beb7d6474aa305761919e0b966abdc9cda3ddbb8302928e9"
163
+ },
164
+ "examples/seedance-action/seedance-2-action-real-input-200px-crf36.mp4": {
165
+ "bytes": 127966,
166
+ "sha256": "503e06db4b6e92a0f658d5acb0ad334c5f251d4d1e5bf22fe1993cfd0822e7ab"
167
+ },
168
+ "examples/seedance-action/seedance-2-action-bicubic-x2-400px.mp4": {
169
+ "bytes": 1091381,
170
+ "sha256": "e09c6ff8d6535ccd0e8e1304e503bc3907251bdd9a71af5bb6139d12bd8ab868"
171
+ },
172
+ "examples/seedance-action/seedance-2-action-clark-x2-400px.mp4": {
173
+ "bytes": 1280429,
174
+ "sha256": "fc1f637292f725a7416479be08ae6eb879ac4584bf8be3fa9d949d521aa8ba68"
175
+ },
176
+ "examples/seedance-action/seedance-2-action-200px-triptych.mp4": {
177
+ "bytes": 5677963,
178
+ "sha256": "2d999796f28d5367789a65fb0393ac58d7f6d0ebe0a7c35197fcc40b4d8804c7"
179
+ },
180
+ "examples/seedance-action/seedance-2-action-200px-before-after.mp4": {
181
+ "bytes": 3661573,
182
+ "sha256": "60f66dc27de71c8ea08866043df3fbc6231cb6e1606587bc5ae5233fdce0ee9d"
183
+ },
184
+ "examples/seedance-action/seedance-2-action-200px-before-after-poster.png": {
185
+ "bytes": 411935,
186
+ "sha256": "57fef502d505a204b2ad06feba3a99c899417850650709c3844342bea739c165"
187
+ },
188
+ "examples/seedance-action/seedance-2-action-200px-wipe.mp4": {
189
+ "bytes": 1978956,
190
+ "sha256": "b24fc83fca9ac2c616c89bcd48f85f98bee819e2fc00917b7206855e45c2b3d6"
191
+ },
192
+ "examples/seedance-action/seedance-2-action-200px-wipe-poster.png": {
193
+ "bytes": 233801,
194
+ "sha256": "91bfb43a0de46fcd5022025ef49f98e90f366332228b7aab4a744e84dd23563f"
195
  }
196
  }
197
  }