Install any skill in seconds. Free to start, no credit card required.
Get Started Free →音频驱动的稀疏帧视频配音工具,支持音频驱动的 Video-to-Video 和 Image-to-Video 生成,实现精准的唇形、头部、身体姿态同步,支持无限时长视频生成
.claude/skills/anbeime-infinitetalk/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 43% | 0% |
| case-02 | ✗→✓ | ▲ Improved | -1% | 0% |
| case-03 | ✗→✓ | ▲ Improved | -5% | 0% |
| case-04 | ✗→✓ | ▲ Improved | 39% | 0% |
| case-05 | ✗→✓ | ▲ Improved | -14% | 0% |
scripts/infer_infinitetalk.py 进行推理input_path: 输入图片路径audio_path: 驱动音频路径(或提供 text 使用 TTS)output_path: 输出视频路径mode: clip(单段)或 streaming(长视频)size: infinitetalk-480(480P)或 infinitetalk-720(720P)sample_steps: 采样步数(默认 40)sample_audio_guide_scale: 音频引导强度(默认 4.0)sample_audio_guide_scale 参数input_path 指向视频文件streaming 模式生成无限时长视频motion_frame 参数控制驱动帧长度(默认 9)(s1) 和 (s2) 区分说话人)size=infinitetalk-480)--quant int8 参数)--offload_model true)bash python scripts/infer_infinitetalk.py \ --input_path ./input.jpg \ --audio_path ./audio.wav \ --output_path ./output.mp4 \ --size infinitetalk-480 \ --mode clip
bash python scripts/infer_infinitetalk.py \ --input_path ./input.jpg \ --audio_path ./long_audio.wav \ --output_path ./long_output.mp4 \ --size infinitetalk-480 \ --mode streaming
bash python scripts/infer_infinitetalk.py \ --input_path ./input.jpg \ --text "你好,今天天气真不错" \ --output_path ./tts_output.mp4 \ --size infinitetalk-480
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 8,332 | 3,634 | -56% | 1 | 1 | 0% | 1,483 | 2,118 | +43% | 0 | 0 | — |
case-02 | fail→pass | 10,368 | 6,657 | -36% | 1 | 1 | 0% | 2,387 | 2,353 | -1% | 0 | 0 | — |
case-03 | fail→pass | 25,072 | 3,480 | -86% | 1 | 1 | 0% | 2,167 | 2,059 | -5% | 0 | 0 | — |
case-04 | fail→pass | 12,163 | 4,954 | -59% | 1 | 1 | 0% | 1,547 | 2,151 | +39% | 0 | 0 | — |
case-05 | fail→pass | 8,752 | 1,412 | -84% | 1 | 1 | 0% | 1,774 | 1,526 | -14% | 0 | 0 | — |
case-06 | pass→pass | 8,553 | 1,469 | -83% | 1 | 1 | 0% | 1,539 | 1,562 | +1% | 0 | 0 | — |
case-07 | fail→pass | 9,311 | 1,699 | -82% | 1 | 1 | 0% | 1,641 | 1,636 | -0% | 0 | 0 | — |
case-08 | fail→pass | 8,583 | 1,209 | -86% | 1 | 1 | 0% | 1,546 | 1,523 | -1% | 0 | 0 | — |
case-09 | fail→pass | 14,946 | 3,492 | -77% | 1 | 1 | 0% | 2,109 | 2,071 | -2% | 0 | 0 | — |
case-10 | pass→pass | 5,630 | 1,586 | -72% | 1 | 1 | 0% | 989 | 1,557 | +57% | 0 | 0 | — |
case-11 | fail→pass | 8,281 | 4,423 | -47% | 1 | 1 | 0% | 1,574 | 2,139 | +36% | 0 | 0 | — |
case-12 | fail→pass | 12,165 | 2,303 | -81% | 1 | 1 | 0% | 1,615 | 1,675 | +4% | 0 | 0 | — |
case-13 | fail→pass | 11,318 | 3,526 | -69% | 1 | 1 | 0% | 2,100 | 1,831 | -13% | 0 | 0 | — |
case-14 | fail→pass | 12,353 | 2,162 | -82% | 1 | 1 | 0% | 2,256 | 1,620 | -28% | 0 | 0 | — |
case-15 | fail→pass | 5,337 | 2,019 | -62% | 1 | 1 | 0% | 1,050 | 1,634 | +56% | 0 | 0 | — |
case-16 | pass→pass | 8,501 | 5,397 | -37% | 1 | 1 | 0% | 1,549 | 2,397 | +55% | 0 | 0 | — |
case-17 | pass→pass | 10,378 | 1,456 | -86% | 1 | 1 | 0% | 1,403 | 1,546 | +10% | 0 | 0 | — |
case-18 | pass→pass | 12,907 | 2,771 | -79% | 1 | 1 | 0% | 1,659 | 1,861 | +12% | 0 | 0 | — |
case-19 | pass→pass | 7,793 | 2,912 | -63% | 1 | 1 | 0% | 1,311 | 1,808 | +38% | 0 | 0 | — |
case-20 | fail→fail | 9,314 | 6,045 | -35% | 1 | 1 | 0% | 1,348 | 2,528 | +88% | 0 | 0 | — |
case-21 | fail→pass | 6,511 | 6,342 | -3% | 1 | 1 | 0% | 1,240 | 2,655 | +114% | 0 | 0 | — |
case-22 | fail→pass | 12,817 | 7,390 | -42% | 1 | 1 | 0% | 2,366 | 2,712 | +15% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted. The headline lift of +68 percentage points is the difference between those two pass rates over the 22 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.