File size: 5,036 Bytes
44e8883
 
 
 
 
 
 
06dff65
 
44e8883
 
06dff65
44e8883
 
 
 
 
 
 
 
06dff65
 
 
 
44e8883
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
---
license: other
pretty_name: StreamTalk Inference Data
language:
  - en
tags:
  - motion-generation
  - co-speech-gesture-generation
  - streaming-generation
  - beat2
  - smpl-x
  - arxiv:2608.01643
size_categories:
  - n<1K
---

# StreamTalk Inference Data

Ready-to-compare SMPL-X predictions from the official retrained StreamTalk CFG checkpoints. This release follows the convenient inference-data style of [SemTalk](https://github.com/Xiangyue-Zhang/SemTalk): users can download generated test outputs and evaluate or visualize them without rerunning WavLM, retrieval, diffusion sampling, and SMPL-X conversion.

- [Paper](https://arxiv.org/abs/2608.01643) · [Hugging Face Paper page](https://huggingface.co/papers/2608.01643)
- [Project page](https://xiangyuezhang.com/StreamTalk/) · [Code](https://github.com/Xiangyue-Zhang/StreamTalk)
- [Model checkpoints](https://huggingface.co/X-Zhang/StreamTalk)

No BEAT2 ground-truth motion, audio, training cache, or model checkpoint is included in these archives.

## Downloads

| Scope | Clips | Checkpoint | CFG | Archive | SHA-256 |
| :-- | --: | :-- | --: | :-- | :-- |
| Speaker 2 (Scott) | 15 | epoch 946 | 3 | [download](https://huggingface.co/datasets/X-Zhang/StreamTalk-Inference-Data/resolve/main/streamtalk_speaker2_e0946_cfg3_npz.tar.gz) | `859cb968c04804a797e3f8eb17aeba1f91bda29305f970564c9be481e41fa452` |
| All speakers | 265 | epoch 940 | 3 | [download](https://huggingface.co/datasets/X-Zhang/StreamTalk-Inference-Data/resolve/main/streamtalk_all_e0940_cfg3_npz.tar.gz) | `295d933883dc40acb9df5f2d2ba02767ec62c9d83f423a78900932499fc8edb4` |

```bash
wget https://huggingface.co/datasets/X-Zhang/StreamTalk-Inference-Data/resolve/main/streamtalk_speaker2_e0946_cfg3_npz.tar.gz
wget https://huggingface.co/datasets/X-Zhang/StreamTalk-Inference-Data/resolve/main/streamtalk_all_e0940_cfg3_npz.tar.gz
sha256sum -c SHA256SUMS

tar -xzf streamtalk_speaker2_e0946_cfg3_npz.tar.gz
tar -xzf streamtalk_all_e0940_cfg3_npz.tar.gz
```

## Released FGD re-score

The downloadable NPZ bytes were independently re-scored with the StreamTalk release scorer and the pinned EMAGE/PantoMatrix AESK encoder (`SHA-256 5cd9566b...c40f55`). Equal-length metric batching uses batch size 16.

| Scope | Archive re-score FGD ↓ | Selection-time FGD ↓ | Absolute difference |
| :-- | --: | --: | --: |
| Speaker 2 (Scott) | **0.3788789702354345** | 0.3788789702354416 | 7.1e-15 |
| All speakers | **0.21764184426140076** | 0.21767150775461275 | 2.97e-5 |

Both differences are below the selected engineering reporting tolerance of `1e-3`. These are accelerated EMAGE/AESK measurements, not a claim of paper-exact B=1/M=1/full-window bitwise equivalence. See `speaker2_fgd.json` and `all_fgd.json` for the complete metric contract.

## NPZ schema

Every archive contains one `res_<BEAT2-test-id>.npz` per official test ID plus a scope manifest. Load files with `allow_pickle=False`.

| Key | Shape/type | Meaning |
| :-- | :-- | :-- |
| `poses` | `[T, 165]`, float32 | local SMPL-X axis-angle rotations |
| `trans` | `[T, 3]`, float32 | root translation |
| `expressions` | `[T, 100]`, float32 | facial expression coefficients |
| `betas` | `[300]`, float32 | body-shape coefficients |
| `gender` | string | body-model metadata |
| `mocap_frame_rate` | scalar | 30 FPS |
| `model` | string | `smplx2020` |

```python
import numpy as np

with np.load("res_2_scott_0_1_1.npz", allow_pickle=False) as motion:
    poses = motion["poses"]
    expressions = motion["expressions"]
    trans = motion["trans"]
```

## Re-score with StreamTalk

The generated files do not include BEAT2 ground truth. After obtaining BEAT2 under its own terms, use the scorer in the [StreamTalk repository](https://github.com/Xiangyue-Zhang/StreamTalk):

```bash
python tools/evaluate_generated_fgd.py \
  --predictions streamtalk_speaker2_e0946_cfg3_npz \
  --ground_truth /path/to/beat_english_v2.0.0/smplxflame_30 \
  --scope speaker2 --device cuda:0 --metric_batch_size 16

python tools/evaluate_generated_fgd.py \
  --predictions streamtalk_all_e0940_cfg3_npz \
  --ground_truth /path/to/beat_english_v2.0.0/smplxflame_30 \
  --scope all --device cuda:0 --metric_batch_size 16
```

## Provenance and terms

- Generator code: StreamTalk training/evaluation commit `b6cac99d5bae5164a21f5db25afef72c3c30556a`.
- Speaker 2 checkpoint SHA-256: `07065d52e333cc48960e4e8fad5c0331f6b9e482f14aa1b78cb721904da5b075`.
- All-speaker checkpoint SHA-256: `3a0e3558d788fd3f11a5a87a8198b42725104b5b9b24f95a4e7c84c14395c9a5`.
- 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.

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/). This dataset card does not redistribute BEAT2 ground truth, audio, or SMPL-X model files.