Home / All Skills / Content & Media Creation / ai-avatar-video
English (EN)简体中文 (ZH)日本語 (JA)
Content & Media Creation · 101-skills/superpowers

ai-avatar-video

Create AI avatar and talking head videos via inference.sh CLI. Recommended: P-Video-Avatar (fastest, cheapest, built-in TTS). Also: OmniHuman, Fabric, PixVerse. Audio: Inworld TTS-2 (100+ languages, emotion steering for characters), ElevenLabs, Kokoro. Capabilities: audio-driven avatars, text-to-avatar, lipsync videos, talking head generation, virtual presenters, UGC content. Use for: AI presenters, explainer videos, virtual influencers, dubbing, marketing videos, UGC ads, gaming avatars, NPC dialogue. Triggers: ai avatar, talking head, lipsync, avatar video, virtual presenter, ai spokesperson, audio driven video, heygen alternative, synthesia alternative, talking avatar, lip sync, video avatar, ai presenter, digital human, ugc, ugc video, ugc ad, avatar ugc

L4 High-Privilege · Graded auth & full auditV1 Commit-PinnedExecute CommandNetwork AccessWrite FileDual-pass Audit
Author 101-skillsGitHub 101-skills/superpowers ↗License MIT（仓库 README 徽标 'License: MIT' 与 .claude-plugin/plugin.json 中 "license": "MIT" 声明；但 pin 树内无 LICENSE 文件，GitHub API 对两 fork 亦报无 license —— 授权只出现在元数据/徽标，无正式许可文本）Commit becc256497
Control levels describe skill capabilities and suggested control measures; they do not represent quality or safety certifications, nor do they dictate usability. Usage decisions rest entirely with users and administrators. Different agent hosts apply varying constraints; do not assume hosts provide automatic protection. Third-party skill copyrights belong to their original authors. Use at your own risk.

1How It Works & Mechanism

Isomorphic to same-repo tool skills: belt CLI whitelist + catalog table; but this skill additionally embeds a 'strong recommendation decision'—marking pruna/p-video-avatar as the top choice and providing a speed/cost comparison table (~1.83s/s, $0.025/sec vs. OmniHuman ~28s/s, $0.16), guiding the agent to default to the most economical path.

tools/video/ai-avatar-video/SKILL.md
| **P-Video-Avatar** | `pruna/p-video-avatar` | **Best overall: speed, cost, quality, control** | **Yes (30 voices, 10 languages)** |
The 'Start with P-Video-Avatar' section + Cost & Speed Comparison table translates 'model selection' into concrete economic metrics, sparing the model from having to compare prices on its own.

Two technical paths depend on whether the model has built-in TTS: models with built-in TTS (p-video-avatar/fabric) directly take image+voice_script; models without TTS (OmniHuman/PixVerse) first run inworld/text-to-speech-2 to obtain an audio_url, then combine image_url+audio_url—the skill strings the chain into reproducible commands via 'Full Workflow'.

tools/video/ai-avatar-video/SKILL.md
belt app run inworld/text-to-speech-2 --input '{
Step 1 uses `> speech.json` redirection to persist TTS output to disk and extract the audio-url, which is then fed into the avatar model in step 2—this two-step assembly demonstrates how examples serve as pipelines.

The dubbing/multilingual pipeline is a 4-stage process: fast-whisper-large-v3 transcription → LLM/human translation → kokoro-tts generating new language audio → latentsync-1-6 for lip-syncing, all executed via belt invocations of infsh/* apps.

tools/video/ai-avatar-video/SKILL.md
belt app run infsh/fast-whisper-large-v3 --input '{"audio_url": "https://video.mp4"}' > transcript.json
'Localizing a video' is decoupled into 4 independently replaceable app steps, each producing a URL/JSON output, naturally allowing for human-in-the-loop review.

UGC batch production is explicitly demonstrated: using a shell for loop to batch-run p-video-avatar with 3 different voices on the same portrait, showing that this skill's positioning encompasses 'scaled production' (and thus directly ties into costs and portrait usage boundaries).

tools/video/ai-avatar-video/SKILL.md
for voice in "Zephyr (Female)" "Puck (Male)" "Aoede (Female)"; do
The batch production example uses voice_script/voice_prompt/video_prompt as loop variable templates—agents can copy them directly into multi-line Bash scripts.

Input asset chain connects with image skills: using pruna/p-image to first generate a 9:16 portrait before creating an avatar (Generate Portrait + Avatar workflow), maintaining a vertical aspect ratio throughout (Tips: output aspect ratio matches the input image).

tools/video/ai-avatar-video/SKILL.md
"prompt": "professional headshot portrait of a young woman, neutral background, looking at camera, studio lighting, photorealistic",
Avatar quality heavily depends on the input portrait (Tips like front-facing, good lighting, etc.), and the skill explicitly incorporates asset specifications.

2Core Capabilities

01Text-driven talking avatars (p-video-avatar with built-in TTS, 30 voices/10 languages, ready to generate with voice_script+voice)
02Audio-driven lip synchronization (feeding image_url+audio_url to OmniHuman/Fabric/PixVerse)
03Multi-character avatars (designating the driven character in multi-person images with OmniHuman 1.5)
04Dubbing/multilingual localization pipeline (whisper → translation → kokoro → latentsync)
05Batch generation of UGC-style content (for loop example with multiple presenters for the same product)
06Style control: voice_prompt controls tone and emotion, video_prompt controls body language and background

3External Dependencies

KindDependency / Endpoint
clibelt（inference.sh CLI）
packagebelt-sh/cli（npx skills add 安装）
api平台头像/TTS/音频 app：pruna/p-video-avatar、pruna/p-image、bytedance/omnihuman-1-5、falai/fabric-1-0、falai/pixverse-lipsync、inworld/text-to-speech-2、infsh/kokoro-tts、infsh/fast-whisper-large-v3、infsh/latentsync-1-6
network平台文档
network安装说明托管

4Recommended Governance & Evidence

Suggested Controls for Safe Operation
For Agent Runtime
Graded/per-use authorization, sandbox recommended
For Administrators
Approval workflow, rollback support, full audit
Evidence & Triggers (File Paths & Excerpts)
  • 网络端点: Declared external network access: platform belt CLI → inference.sh platform (avatar generation/TTS/STT/lip-sync apps) ['inference.sh',
  • 本地工作目录: File I/O and artifact output: the multi-step workflow in examples uses shell redirection to persist intermediate artifacts: `> speech.json` (TTS output), `> transcript.json` (transcription), `
  • SKILL.md: Declared invocation of host CLI interpreter or external runtime environment
Audit Notes: Structured identically to in-repo tool skills: no script files, no credential reads, no direct network calls; all actions are consolidated into the belt CLI, targeting predictable external endpoints. Additional considerations: 1) Targets involve real human portraits/voices, where mass-producing UGC spokesperson content touches on portrait rights, voice rights, and deepfake misuse boundaries; it is recommended to use only owned/licensed assets and disclose AI generation. 2) Billing is per second per generation (starting at $0.025/s), so batch production with for loops directly scales up costs. 3) Supply chain risks as noted previously (unpinned npx, inference-sh mirror, no LICENSE file). Per standard grading guidelines: the `belt` CLI called by this skill is distributed via third-party npm/npx (`npx skills add belt-sh/cli`), runs under authenticated session (`belt login`), and consumes inference.sh account quota/credits—falling under 'calls third-party CLI with quota/auth state', thus assigned Orange (use after evaluation); the skill source code itself contains no token/key reading.

5Second-Pass Verification

  • [discrepancy] description 声称 Inworld TTS-2 支持 100+ languages、emotion steering — 正文有 inworld/text-to-speech-2 调用示例且含 [friendly]/[excited] 情绪标记（emotion steering 有据）；但 '100+ languages' 这一具体数字在 SKILL.md 内无出处——数值来自平台宣称，本源码无法定位。
  • [ok] 成本对比表自洽性 — 表内 ~28s/s ≈ 1.83s/s×15（'15x slower' 一致）、$0.16≈$0.025×6.4（'6.4x more' 一致）；数字为平台方宣称，作为选型参考存在。
  • [ok] P-Video-Avatar 30 voices/10 languages 声明 — 表内 Built-in TTS 列与正文 UGC 章节（'30 voices, 10 languages — match your target audience'）两处一致。
  • [ok] 四条工作流（TTS+Avatar / 配音流水线 / 画像→头像 / UGC 批产）可执行性 — 每条均含完整 belt 命令与重定向；调用的 app-id（inworld/text-to-speech-2、infsh/fast-whisper-large-v3、infsh/kokoro-tts、infsh/latentsync-1-6、pruna/p-image）均在示例内成对出现。
  • [ok] 无脚本文件/无凭证/无直连网络 — 目录仅 SKILL.md；执行面为 allowed-tools 白名单 belt；多行 shell 仅以示例文本存在，由宿主执行。
  • [ok] 本地文件写入面 — 仅示例重定向 > speech.json / > transcript.json / > new_speech.json，属执行示例时的落盘行为。
  • [ok] 肖像/音频为远程 URL 输入 — 全部示例输入 image/audio/image_url/audio_url 值为 https://…，无本地路径。
  • [unlocatable] installs 安装量 — 无公开安装量数据源，置 null。

6Verdict & Scenarios

  • Decision-making shifted maximally upfront: recommended model + cost/speed comparison table + tips on when to use alternative paths, allowing mass-production users to execute immediately.
  • Turnkey multilingual localization pipeline: four steps (whisper → kokoro → latentsync) enable dubbing an existing video into a new language.
  • Dual-track style control (voice_prompt for tone + video_prompt for body/background) clearly delineated in examples, offering high output tunability.
  • Complete end-to-end chaining within the same ecosystem: portrait generation (p-image) → avatar (p-video-avatar) → downstream video/publishing workflows (complete interlinking across Related Skills).
    • Best for: Suitable for marketing teams with authorized assets (owned/licensed portrait rights) conducting mass production of AI spokespersons, UGC-style ads, course presentations, and multilingual localization; also suitable for users wanting to run low-cost test videos with p-video-avatar before deciding whether to adopt OmniHuman/PixVerse.
    Not for: Not suitable for scenarios lacking portrait/voice authorization or handling real individuals' likenesses without consent; not suitable for scenarios with zero tolerance for outbound asset transfer and retention; not suitable for users seeking deep, fine-grained tuning of a single model (specialized standalone skills like dedicated OmniHuman packages should be used instead); not suitable for automated batch tasks with no cost budget.
    Install Agent-Ready
    ① Verified Mirror 2026-09-09
    Option A · Direct DownloadDownload ai-avatar-video.tar.gz
    sha256: 28ad95bf92e9736e…
    Option B · Copy Guide for Agent
    install.json
    Agents automatically download and verify mirrors from this JSON guide.
    ② Upstream GitHub
    GitHub Repo ↗
    Pinned to commit becc256497.