Datasets:
Languages:
English
Size:
n<1K
ArXiv:
Tags:
motion-generation
co-speech-gesture-generation
streaming-generation
beat2
smpl-x
arxiv:2608.01643
License:
Publish clip-matched BEAT2 predictions and ground truth
Browse files- MANIFEST.json +30 -9
- README.md +12 -5
- SHA256SUMS +4 -2
- StreamTalk_25speakers_predictions_and_gt.zip +3 -0
- StreamTalk_speaker2_predictions_and_gt.zip +3 -0
MANIFEST.json
CHANGED
|
@@ -7,17 +7,17 @@
|
|
| 7 |
"archive": "streamtalk_all_e0940_cfg3_npz.tar.gz",
|
| 8 |
"archive_sha256": "295d933883dc40acb9df5f2d2ba02767ec62c9d83f423a78900932499fc8edb4",
|
| 9 |
"archive_size_bytes": 552982864,
|
| 10 |
-
"cfg_scale": 3
|
| 11 |
"checkpoint": "model.pt",
|
| 12 |
"checkpoint_sha256": "3a0e3558d788fd3f11a5a87a8198b42725104b5b9b24f95a4e7c84c14395c9a5",
|
| 13 |
"clip_batch_size": 16,
|
| 14 |
"clip_count": 265,
|
| 15 |
"format": "streamtalk-inference-scope-v1",
|
| 16 |
"generation_seconds_by_rank": [
|
| 17 |
-
0
|
| 18 |
-
0
|
| 19 |
-
0
|
| 20 |
-
0
|
| 21 |
],
|
| 22 |
"generation_note": "The successful all-speaker outputs were schema/hash verified and reused during final packaging; the resumed packaging pass did not repeat generation timing.",
|
| 23 |
"published_artifact_rescore": {
|
|
@@ -31,13 +31,23 @@
|
|
| 31 |
"metric_batch_size": 16,
|
| 32 |
"retrieval_query_frames": 8,
|
| 33 |
"scope": "all",
|
| 34 |
-
"seed": 0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 35 |
},
|
| 36 |
"speaker2": {
|
| 37 |
"archive": "streamtalk_speaker2_e0946_cfg3_npz.tar.gz",
|
| 38 |
"archive_sha256": "859cb968c04804a797e3f8eb17aeba1f91bda29305f970564c9be481e41fa452",
|
| 39 |
"archive_size_bytes": 29238804,
|
| 40 |
-
"cfg_scale": 3
|
| 41 |
"checkpoint": "model.pt",
|
| 42 |
"checkpoint_sha256": "07065d52e333cc48960e4e8fad5c0331f6b9e482f14aa1b78cb721904da5b075",
|
| 43 |
"clip_batch_size": 1,
|
|
@@ -60,8 +70,19 @@
|
|
| 60 |
"metric_batch_size": 16,
|
| 61 |
"retrieval_query_frames": 8,
|
| 62 |
"scope": "speaker2",
|
| 63 |
-
"seed": 0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 64 |
}
|
| 65 |
},
|
| 66 |
-
"seed": 0
|
|
|
|
| 67 |
}
|
|
|
|
| 7 |
"archive": "streamtalk_all_e0940_cfg3_npz.tar.gz",
|
| 8 |
"archive_sha256": "295d933883dc40acb9df5f2d2ba02767ec62c9d83f423a78900932499fc8edb4",
|
| 9 |
"archive_size_bytes": 552982864,
|
| 10 |
+
"cfg_scale": 3,
|
| 11 |
"checkpoint": "model.pt",
|
| 12 |
"checkpoint_sha256": "3a0e3558d788fd3f11a5a87a8198b42725104b5b9b24f95a4e7c84c14395c9a5",
|
| 13 |
"clip_batch_size": 16,
|
| 14 |
"clip_count": 265,
|
| 15 |
"format": "streamtalk-inference-scope-v1",
|
| 16 |
"generation_seconds_by_rank": [
|
| 17 |
+
0,
|
| 18 |
+
0,
|
| 19 |
+
0,
|
| 20 |
+
0
|
| 21 |
],
|
| 22 |
"generation_note": "The successful all-speaker outputs were schema/hash verified and reused during final packaging; the resumed packaging pass did not repeat generation timing.",
|
| 23 |
"published_artifact_rescore": {
|
|
|
|
| 31 |
"metric_batch_size": 16,
|
| 32 |
"retrieval_query_frames": 8,
|
| 33 |
"scope": "all",
|
| 34 |
+
"seed": 0,
|
| 35 |
+
"contains_ground_truth": false,
|
| 36 |
+
"paired_release": {
|
| 37 |
+
"archive": "StreamTalk_25speakers_predictions_and_gt.zip",
|
| 38 |
+
"archive_sha256": "e1e7304517e08dea0979d0577d7d86e03527b6f5e57932e384f0c517f3527bda",
|
| 39 |
+
"size_bytes": 1341193151,
|
| 40 |
+
"prediction_count": 265,
|
| 41 |
+
"ground_truth_count": 265,
|
| 42 |
+
"contains_ground_truth": true,
|
| 43 |
+
"ground_truth_source": "https://huggingface.co/datasets/haiyangliu0/BEAT2/tree/8689ecb43513ba31964fd60e0ca69be02d3b0872/beat_english_v2.0.0/smplxflame_30"
|
| 44 |
+
}
|
| 45 |
},
|
| 46 |
"speaker2": {
|
| 47 |
"archive": "streamtalk_speaker2_e0946_cfg3_npz.tar.gz",
|
| 48 |
"archive_sha256": "859cb968c04804a797e3f8eb17aeba1f91bda29305f970564c9be481e41fa452",
|
| 49 |
"archive_size_bytes": 29238804,
|
| 50 |
+
"cfg_scale": 3,
|
| 51 |
"checkpoint": "model.pt",
|
| 52 |
"checkpoint_sha256": "07065d52e333cc48960e4e8fad5c0331f6b9e482f14aa1b78cb721904da5b075",
|
| 53 |
"clip_batch_size": 1,
|
|
|
|
| 70 |
"metric_batch_size": 16,
|
| 71 |
"retrieval_query_frames": 8,
|
| 72 |
"scope": "speaker2",
|
| 73 |
+
"seed": 0,
|
| 74 |
+
"contains_ground_truth": false,
|
| 75 |
+
"paired_release": {
|
| 76 |
+
"archive": "StreamTalk_speaker2_predictions_and_gt.zip",
|
| 77 |
+
"archive_sha256": "05aa172ef8ccc8c92c9aee6e921360e287972e17d8c07654ebcd8f8a64bf9f2a",
|
| 78 |
+
"size_bytes": 71001609,
|
| 79 |
+
"prediction_count": 15,
|
| 80 |
+
"ground_truth_count": 15,
|
| 81 |
+
"contains_ground_truth": true,
|
| 82 |
+
"ground_truth_source": "https://huggingface.co/datasets/haiyangliu0/BEAT2/tree/8689ecb43513ba31964fd60e0ca69be02d3b0872/beat_english_v2.0.0/smplxflame_30"
|
| 83 |
+
}
|
| 84 |
}
|
| 85 |
},
|
| 86 |
+
"seed": 0,
|
| 87 |
+
"contains_ground_truth": true
|
| 88 |
}
|
README.md
CHANGED
|
@@ -22,9 +22,16 @@ Ready-to-compare SMPL-X predictions from the official retrained StreamTalk CFG c
|
|
| 22 |
- [Project page](https://xiangyuezhang.com/StreamTalk/) · [Code](https://github.com/Xiangyue-Zhang/StreamTalk)
|
| 23 |
- [Model checkpoints](https://huggingface.co/X-Zhang/StreamTalk)
|
| 24 |
|
| 25 |
-
|
| 26 |
|
| 27 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 28 |
|
| 29 |
| Scope | Clips | Checkpoint | CFG | Archive | SHA-256 |
|
| 30 |
| :-- | --: | :-- | --: | :-- | :-- |
|
|
@@ -53,7 +60,7 @@ Both differences are below the selected engineering reporting tolerance of `1e-3
|
|
| 53 |
|
| 54 |
## NPZ schema
|
| 55 |
|
| 56 |
-
|
| 57 |
|
| 58 |
| Key | Shape/type | Meaning |
|
| 59 |
| :-- | :-- | :-- |
|
|
@@ -76,7 +83,7 @@ with np.load("res_2_scott_0_1_1.npz", allow_pickle=False) as motion:
|
|
| 76 |
|
| 77 |
## Re-score with StreamTalk
|
| 78 |
|
| 79 |
-
|
| 80 |
|
| 81 |
```bash
|
| 82 |
python tools/evaluate_generated_fgd.py \
|
|
@@ -97,4 +104,4 @@ python tools/evaluate_generated_fgd.py \
|
|
| 97 |
- All-speaker checkpoint SHA-256: `3a0e3558d788fd3f11a5a87a8198b42725104b5b9b24f95a4e7c84c14395c9a5`.
|
| 98 |
- Seed: 0. Speaker 2 generation uses clip batch 1; all-speaker generation uses clip batch 16; both use CFG 3 and an 8-frame retrieval query.
|
| 99 |
|
| 100 |
-
The predictions are derived from experiments on BEAT2 and use the SMPL-X body model. Users remain responsible for complying with the licenses and terms of the [BEAT2 dataset](https://huggingface.co/datasets/H-Liu1997/BEAT2) and [SMPL-X](https://smpl-x.is.tue.mpg.de/).
|
|
|
|
| 22 |
- [Project page](https://xiangyuezhang.com/StreamTalk/) · [Code](https://github.com/Xiangyue-Zhang/StreamTalk)
|
| 23 |
- [Model checkpoints](https://huggingface.co/X-Zhang/StreamTalk)
|
| 24 |
|
| 25 |
+
## Compare with ground truth
|
| 26 |
|
| 27 |
+
| BEAT2 evaluation protocol | Predictions | Ground-truth clips | Download |
|
| 28 |
+
| --- | ---: | ---: | --- |
|
| 29 |
+
| Speaker 2 (Scott) | 15 | 15 | [Speaker 2 predictions and GT](https://huggingface.co/datasets/X-Zhang/StreamTalk-Inference-Data/resolve/main/StreamTalk_speaker2_predictions_and_gt.zip) |
|
| 30 |
+
| 25 English speakers | 265 | 265 | [25-speaker predictions and GT](https://huggingface.co/datasets/X-Zhang/StreamTalk-Inference-Data/resolve/main/StreamTalk_25speakers_predictions_and_gt.zip) |
|
| 31 |
+
|
| 32 |
+
Each paired ZIP contains the original prediction files under `predictions/res_<id>.npz` and the unmodified [BEAT2 reference motions](https://huggingface.co/datasets/haiyangliu0/BEAT2/tree/8689ecb43513ba31964fd60e0ca69be02d3b0872/beat_english_v2.0.0/smplxflame_30) under `ground_truth/<id>.npz`. The release scorer reads these directories directly. The file `pairs.csv` records both frame counts and SHA-256 hashes for every clip. These are the same predicted bytes used in the released re-score below. [ECCV proceedings](https://doi.org/10.1007/978-3-032-37595-7_32).
|
| 33 |
+
|
| 34 |
+
## Original prediction-only downloads
|
| 35 |
|
| 36 |
| Scope | Clips | Checkpoint | CFG | Archive | SHA-256 |
|
| 37 |
| :-- | --: | :-- | --: | :-- | :-- |
|
|
|
|
| 60 |
|
| 61 |
## NPZ schema
|
| 62 |
|
| 63 |
+
Each original prediction-only archive contains one `res_<BEAT2-test-id>.npz` per official test ID plus a scope manifest. Load files with `allow_pickle=False`.
|
| 64 |
|
| 65 |
| Key | Shape/type | Meaning |
|
| 66 |
| :-- | :-- | :-- |
|
|
|
|
| 83 |
|
| 84 |
## Re-score with StreamTalk
|
| 85 |
|
| 86 |
+
Extract a paired ZIP, then pass its `predictions` and `ground_truth` directories to the scorer in the [StreamTalk repository](https://github.com/Xiangyue-Zhang/StreamTalk). For the original prediction-only TAR files, supply the corresponding reference directory from BEAT2.
|
| 87 |
|
| 88 |
```bash
|
| 89 |
python tools/evaluate_generated_fgd.py \
|
|
|
|
| 104 |
- All-speaker checkpoint SHA-256: `3a0e3558d788fd3f11a5a87a8198b42725104b5b9b24f95a4e7c84c14395c9a5`.
|
| 105 |
- Seed: 0. Speaker 2 generation uses clip batch 1; all-speaker generation uses clip batch 16; both use CFG 3 and an 8-frame retrieval query.
|
| 106 |
|
| 107 |
+
The predictions are derived from experiments on BEAT2 and use the SMPL-X body model. Users remain responsible for complying with the licenses and terms of the [BEAT2 dataset](https://huggingface.co/datasets/H-Liu1997/BEAT2) and [SMPL-X](https://smpl-x.is.tue.mpg.de/). The paired ZIPs include the original BEAT2 motion references. The two original prediction-only TAR files each contain zero ground-truth clips.
|
SHA256SUMS
CHANGED
|
@@ -1,8 +1,10 @@
|
|
| 1 |
-
|
| 2 |
-
|
| 3 |
15d35f63d787752ab852e849df7369166567cc951429007674e5a658de3a6a9f speaker2_fgd.json
|
| 4 |
9a859040981909ead510e5bd34e9d066b9d87aba7e8471583655a57bc10459b6 all_fgd.json
|
| 5 |
d4031ae4e6180fd4eb366de3e0c5c4c849046c9cac79963d7293ab3a432321f3 speaker2/MANIFEST.json
|
| 6 |
b42c459ec32a71a97a1fbbd86b7a09f80f00a9786e5783f9a8da19e3bf845878 all_speakers/MANIFEST.json
|
| 7 |
859cb968c04804a797e3f8eb17aeba1f91bda29305f970564c9be481e41fa452 streamtalk_speaker2_e0946_cfg3_npz.tar.gz
|
| 8 |
295d933883dc40acb9df5f2d2ba02767ec62c9d83f423a78900932499fc8edb4 streamtalk_all_e0940_cfg3_npz.tar.gz
|
|
|
|
|
|
|
|
|
| 1 |
+
3ce40a84eb4cf45894b148caa7e425d4097f07983a57bf60e7dd54c031c9f83e README.md
|
| 2 |
+
2ce9c5c4dcdd675a700cad3387598ea1747a36f89b82e3da71208351930ac700 MANIFEST.json
|
| 3 |
15d35f63d787752ab852e849df7369166567cc951429007674e5a658de3a6a9f speaker2_fgd.json
|
| 4 |
9a859040981909ead510e5bd34e9d066b9d87aba7e8471583655a57bc10459b6 all_fgd.json
|
| 5 |
d4031ae4e6180fd4eb366de3e0c5c4c849046c9cac79963d7293ab3a432321f3 speaker2/MANIFEST.json
|
| 6 |
b42c459ec32a71a97a1fbbd86b7a09f80f00a9786e5783f9a8da19e3bf845878 all_speakers/MANIFEST.json
|
| 7 |
859cb968c04804a797e3f8eb17aeba1f91bda29305f970564c9be481e41fa452 streamtalk_speaker2_e0946_cfg3_npz.tar.gz
|
| 8 |
295d933883dc40acb9df5f2d2ba02767ec62c9d83f423a78900932499fc8edb4 streamtalk_all_e0940_cfg3_npz.tar.gz
|
| 9 |
+
05aa172ef8ccc8c92c9aee6e921360e287972e17d8c07654ebcd8f8a64bf9f2a StreamTalk_speaker2_predictions_and_gt.zip
|
| 10 |
+
e1e7304517e08dea0979d0577d7d86e03527b6f5e57932e384f0c517f3527bda StreamTalk_25speakers_predictions_and_gt.zip
|
StreamTalk_25speakers_predictions_and_gt.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:e1e7304517e08dea0979d0577d7d86e03527b6f5e57932e384f0c517f3527bda
|
| 3 |
+
size 1341193151
|
StreamTalk_speaker2_predictions_and_gt.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:05aa172ef8ccc8c92c9aee6e921360e287972e17d8c07654ebcd8f8a64bf9f2a
|
| 3 |
+
size 71001609
|