本文へ移動
cccskills
無料GitHub で公開

byted-byteplus-vod-video-enhancement

Upload video/audio media to BytePlus VOD (Video on Demand) storage, returning the Vid and playback references; supports both local file upload (ApplyUploadInfo + TOS + CommitUploadInfo) and URL pull upload (UploadMediaByUrl); also supports AI-based comprehensive quality restoration on already-uploaded videos (removing compression artifacts, noise, scratches, improving clarity, and configuring Pro-tier color depth). Trigger keywords: upload video, upload media, upload to VOD, URL upload, pull upload, local upload, file upload, UploadMediaByUrl, ApplyUploadInfo, media ingestion, quality restoration, quality enhancement, comprehensive restoration, video denoising, denoise, compression artifact removal, color depth, bit depth, 10-bit, 12-bit, 16-bit.

インストール方法を見る

含まれるファイル(10)

  • SKILL.md14.5 KB
  • LICENSE9.9 KB
  • references/quality-enhance.md6.0 KB
  • scripts/env.md218 B
  • scripts/poll_execution.py1.3 KB
  • scripts/pyproject.toml237 B
  • scripts/quality_enhance.py11.7 KB
  • scripts/tos_upload.py13.5 KB
  • scripts/upload.py12.3 KB
  • scripts/vod_client.py23.3 KB

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

VOD_video enhancement

Uploads video/audio to a BytePlus VOD space (from a local file or a public URL) and returns a vid://vxxxx reference. Additionally provides AI-based comprehensive quality restoration that removes compression artifacts, noise, and scratches from ingested videos, improving overall clarity and color rendition.


Prerequisites

  • Environment variables (required, can be configured via a .env file in the working directory — the scripts will load it automatically):
    • BYTEPLUS_ACCESSKEY — BytePlus Access Key
    • BYTEPLUS_SECRETKEY — BytePlus Secret Key
    • VOD_SPACE_NAME — VOD space name
  • Execution: examples use uv run python ... (if the host environment can run Python directly, python scripts/... also works).

Workflow Overview

Upload pipeline (local file):
  [S1_APPLY]  ApplyUploadInfo → returns TOS upload address + SessionKey
  [S2_TOS]    PUT file to TOS (direct or chunked)
  [S3_COMMIT] CommitUploadInfo → returns Vid
  Output: { Vid, Source, PlayURL, FileName, SpaceName, SourceUrl }

Upload pipeline (URL):
  [S1_UPLOAD] Submit URL upload job (UploadMediaByUrl) → returns JobId
  [S2_POLL]   Poll QueryUploadTaskInfo → returns Vid
  Output: { Vid, Source, PlayURL, FileName, SpaceName, SourceUrl, JobId }

Quality restoration pipeline:
  [S3_ENHANCE] Submit restoration job (StartExecution/enhanceVideo) → returns RunId
  [S4_POLL]    Poll GetExecution → returns the restored file
  Output: { Status, SpaceName, VideoUrls[{ FileId, DirectUrl, Source }] }

Quick Self-Check (recommended)

Before running any script, confirm the following (avoid unrelated Python/uv version checks):

  • .env or environment variables contain:
    • BYTEPLUS_ACCESSKEY + BYTEPLUS_SECRETKEY
    • VOD_SPACE_NAME

Once verified, pick the corresponding pipeline based on user intent:

User intentPipelineEntry script
Upload video to VODUpload pipelinescripts/upload.py
Quality restoration / denoise / remove compression artifactsQuality restoration pipelinescripts/quality_enhance.py

S1_UPLOAD & S2_POLL: Upload and Obtain Vid

Calling Convention

Run from the Skill root directory (byted-byteplus-vod-video-enhancement/):

# Local file upload (synchronous — returns Vid when complete)
uv run python scripts/upload.py "/path/to/video.mp4" [space_name]

# URL upload (automatically polls until a Vid is returned)
uv run python scripts/upload.py "<https://example.com/video.mp4>" [space_name]

# Example: specifying the space
uv run python scripts/upload.py "https://example.com/sample.mp4" my_space
  • First argument: either a local file path or a public http:// / https:// link. The script auto-detects which mode to use.
  • Second argument (optional): the VOD space name; when omitted it is read from the environment variable VOD_SPACE_NAME.
  • The file / URL must carry a file extension (such as .mp4, .mov, .mp3), otherwise an error is raised.

Upload Flow

Local file upload (synchronous, three-step):

  1. Call ApplyUploadInfo (API Version: 2023-01-01) to obtain the TOS upload address, authentication token, and SessionKey.
  2. PUT the file to TOS (direct upload for files < 20 MiB, chunked upload otherwise).
  3. Call CommitUploadInfo (API Version: 2023-01-01) with the SessionKey; returns the Vid.

URL upload (two-phase asynchronous):

  1. Call UploadMediaByUrl (API Version: 2023-01-01) to submit the pull job; returns a JobId.
  2. Poll QueryUploadTaskInfo until the job completes, with a maximum wait of 30 minutes (360 × 5s).
  3. Once the job is complete, return the Vid.

Output Format

On success, a JSON line is printed to stdout:

{
  "Vid": "v0d123abc",
  "Source": "vid://v0d123abc",
  "PlayURL": "https://example.cdn.com/xxx.m3u8",
  "PosterUri": "",
  "FileName": "uuid-filename.mp4",
  "SpaceName": "my_space",
  "SourceUrl": "https://example.com/video.mp4",
  "JobId": "job-xxx"
}
  • Source: a vid://-formatted reference that can be passed directly to follow-up skills such as byted-mediakit.
  • The host agent should save the Source field for use in subsequent processing steps.

Timeout Handling

If polling times out (30 minutes), the output is:

{
  "error": "Polling timed out (360 attempts × 5s); the URL pull upload is still processing",
  "resume_hint": {
    "description": "The URL upload has not finished yet; retry with the command below",
    "command": "uv run python scripts/upload.py \"<original URL>\" [space_name]"
  },
  "JobIds": "job-xxx",
  "State": "running"
}

S3_ENHANCE & S4_POLL: AI Comprehensive Quality Restoration

Calling Convention

Run from the Skill root directory (byted-byteplus-vod-video-enhancement/):

# Submit after the user has explicitly selected both config and repair_style
uv run python scripts/quality_enhance.py '{"type":"Vid","video":"v0310abc","config":"common","repair_style":1}'

# Example: a vid:// prefix is also accepted (the script strips it automatically)
uv run python scripts/quality_enhance.py '{"type":"Vid","video":"vid://v0d225gxxx","config":"common","repair_style":1}' production_space

# Optional target output resolution (omit res for source resolution)
uv run python scripts/quality_enhance.py '{"type":"Vid","video":"v0310abc","config":"common","repair_style":1,"res":"1080p"}'

# Optional color depth (Pro only; online output supports 8, 10, 12, or 16 bit)
uv run python scripts/quality_enhance.py '{"type":"Vid","video":"v0310abc","config":"common","repair_style":2,"bit_depth":10}'

# 16-bit FFV1 output (input video must be 40 seconds or shorter)
uv run python scripts/quality_enhance.py '{"type":"Vid","video":"v0310abc","config":"common","repair_style":2,"bit_depth":16}'

# Pass parameters via @file.json (recommended — avoids shell escaping issues)
uv run python scripts/quality_enhance.py @params.json

# Resume polling after a timeout
uv run python scripts/poll_execution.py '<RunId>' [space_name]

Parameter Reference

ParameterTypeRequiredDescription
typestring✅Vid (video ID) or DirectUrl (VOD storage FileName)
videostring✅The video Vid or FileName (a vid:// prefix is accepted and automatically stripped)
configstring✅VolcMoeEnhanceParam Config; one of common, ugc, short_series, aigc, old_film. If the user explicitly asks for defaults, use common.
repair_styleinteger✅VolcMoeEnhanceParam VideoStrategy.RepairStyle; 1 = Standard, 2 = Pro. If the user explicitly asks for defaults, use 1.
resstringnoOptional MoeEnhance.Target.Res — target output resolution. Omit or leave empty for source resolution (no upscaling target). Allowed: 240p, 360p, 480p, 540p, 720p, 1080p, 2k, 4k.
bit_depthintegernoOptional MoeEnhance.Target.BitDepth; Pro tier only. Online output supports 8, 10, 12, or 16. Omit it to use the API default, 8. 8 uses H.264; 10 and 12 use H.265; 16 uses lossless FFV1.

Before quality restoration, you MUST ask the user to choose both required enhancement parameters if either config or repair_style is missing. Do not silently use defaults. Only use config=common and repair_style=1 when the user explicitly asks for default/recommended settings. When asking the user, use plain product language only; do not show internal parameter names or values such as config=..., repair_style=..., common, or short_series in the question text or option labels.

Suggested prompt:

Video enhancement may take some time. Choosing the right template usually gives better results.
What type of video is it?

  1. General video
  2. Short video / UGC
  3. Short drama / short series
  4. AI-generated content
  5. Old film / classic footage that needs restoration

Which video enhancement tier would you like to use?

  1. Standard: balanced visual improvement and processing speed
  2. Pro: cinematic-grade restoration with longer processing time; allowlist access may be required

What output resolution do you want? (optional)

  • Keep the same as the source video (recommended default — do not pass res)
  • Or choose a target: 240p, 360p, 480p, 540p, 720p, 1080p, 2K, 4K

If you chose Pro, what color depth do you want? (optional)

  • 8-bit (API default, H.264, broad compatibility)
  • 10-bit (recommended for premium delivery and HDR workflows, H.265)
  • 12-bit (professional post-production, H.265; requires compatible professional tools and displays)
  • 16-bit (lossless FFV1; source video must be 40 seconds or shorter; processed serially and may take longer)

If the user wants source / original resolution, omit res from the JSON (or use an empty string). Only set res when they explicitly pick a target resolution.

If the user asks for a default recommendation, use config=common and repair_style=1 with no res or bit_depth. Otherwise, wait for the user's selections before running scripts/quality_enhance.py.

Internal mapping: General video -> config=common; Short video / UGC -> config=ugc; Short drama / short series -> config=short_series; AI-generated content -> config=aigc; Old film / classic footage -> config=old_film; Standard -> repair_style=1; Pro -> repair_style=2; source resolution -> omit res; 240p–4K -> res as listed in the parameter table (2k / 4k in JSON); selected color depth -> bit_depth=8, 10, 12, or 16. Do not expose these parameter names or values in the question unless the user asks for implementation details.

Color-depth usage limits:

  • bit_depth can be submitted only with Pro (repair_style=2) and is part of the advanced enhancement output settings.
  • The online API supports 8-bit, 10-bit, 12-bit, and 16-bit output.
  • Before submitting 16-bit, confirm that the input video is 40 seconds or shorter. The script cannot infer source duration from the current arguments, so do not submit until this requirement is confirmed.
  • 16-bit uses FFV1 lossless encoding, so do not configure a target bitrate. This skill does not expose or send a target bitrate.
  • 16-bit tasks are processed serially, one at a time. Set the expectation that processing may take longer and wait for the current task to finish before submitting another 16-bit task.
  • 10-bit and 12-bit output uses H.265. For 12-bit review, ordinary players and SDR operating-system pipelines may render incorrectly; recommend a color-managed professional tool and monitor. The documentation recommends MOV with H.265 yuv444p12le for 12-bit delivery because conventional MP4/H.264/H.265 workflows are generally limited to 8-bit or 10-bit and may dither or truncate higher precision.
  • Ordinary players and SDR rendering pipelines may also render 16-bit incorrectly. Review 16-bit output in a color-managed professional tool on a compatible monitor, and use high-throughput storage such as a dedicated NVMe SSD for large lossless assets.

Special handling for Pro: if the user chooses repair_style=2 and the StartExecution/GetExecution response returns HTTP status 403, or any error message contains Permission denied, explain that Pro is only available to users on the allowlist. Ask the user to submit a ticket to apply: https://console.byteplus.com/workorder/create

Output Format

On success, a JSON line is printed to stdout:

{
  "Status": "Success",
  "SpaceName": "my_space",
  "VideoUrls": [
    {
      "FileId": "xxx",
      "DirectUrl": "path/to/output.mp4",
      "Source": "directurl://path/to/output.mp4",
      "Url": "https://example.cdn.com/path/to/output.mp4?auth_key=..."
    }
  ],
  "AudioUrls": [],
  "Texts": []
}
  • VideoUrls[0].Url: a directly accessible/downloadable URL (the script signs it based on the space's domain/auth rules).
  • VideoUrls[0].Source (directurl://...) can be passed directly to downstream skills.

Timeout Handling

If polling times out (30 minutes), the output is:

{
  "error": "Polling timed out (360 attempts × 5s); the job is still processing",
  "resume_hint": {
    "description": "The job has not finished yet; resume polling with the command below",
    "command": "uv run python scripts/poll_execution.py '<RunId>' [space_name]"
  }
}

Environment Variables

NameDescriptionRequired
BYTEPLUS_ACCESSKEYBytePlus Access KeyYes
BYTEPLUS_SECRETKEYBytePlus Secret KeyYes
VOD_SPACE_NAMEVOD space nameYes (or via CLI argument)
VOD_POLL_INTERVALPolling interval (seconds, default 5)No
VOD_POLL_MAXMaximum polling attempts (default 360)No
VOD_URL_EXPIRE_MINUTESSigned URL expiration (minutes, default 60)No
VOD_PLAY_DOMAINForce the use of a specific play domain (optional, highest priority)No

Error Output Format

All errors share the same format:

{"error": "error description"}

References

  • BytePlus VOD Python SDK
  • Quality restoration parameter reference
  • API: ApplyUploadInfo (Version: 2023-01-01)
  • API: CommitUploadInfo (Version: 2023-01-01)
  • API: UploadMediaByUrl (Version: 2023-01-01)
  • API: QueryUploadTaskInfo (Version: 2023-01-01)
  • API: StartExecution (Version: 2025-07-01)
  • API: GetExecution (Version: 2025-07-01)

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

Deploy, update, validate, and troubleshoot the AgentKit hybrid-cloud customer-service demo in this directory. Use when a user asks an AI coding agent to follow the README, deploy or update the demo, configure OpenAPI/Runtime/Knowledge/Memory/Sandbox/MCP/Skills/A2A, run customer-view validation, or record manual platform steps and failures.

日本語の概要は準備中です。原文の説明を表示しています。

bytedance/agentkit-samples4712026年10月12日 更新

Generate deterministic SVG algorithmic artwork. Invoke when the user asks for geometric, generative, or algorithmic visual art.

日本語の概要は準備中です。原文の説明を表示しています。

bytedance/agentkit-samples4712026年10月12日 更新

通过本地 Python CLI 和 OpenAPI 客户端管理、排查火山云手机资源。适用于查询实例和资源、截图、执行命令、查看任务、检查应用、主机和机房容量、标签、DNS、路由,以及操作已授权的测试云手机实例。

日本語の概要は準備中です。原文の説明を表示しています。

bytedance/agentkit-samples4712026年10月12日 更新

Launch and continue the Volcengine AI Research survey workflow for concept testing, audience design, questionnaire drafting, interview guide generation, execution confirmation, progress checks, and result queries. Use this skill when the user wants to create, revise, confirm, execute, or follow up on a real AI research survey task in ABCompass instead of doing generic brainstorming, copywriting, translation, summarization, or broad market discussion.

日本語の概要は準備中です。原文の説明を表示しています。

bytedance/agentkit-samples4712026年10月12日 更新

Create and check long-running video material evaluation tasks. Use this skill when the user wants to submit videos for evaluation, check an existing video evaluation task list, or fetch the result of a previously created video evaluation task. This skill is not a general-purpose video upload skill because upload is allowed only as an internal step of task creation. Authentication uses an API key passed as an Authorization bearer token.

日本語の概要は準備中です。原文の説明を表示しています。

bytedance/agentkit-samples4712026年10月12日 更新

火山引擎 AntiDDoSPro 高防域名只读巡检。用户要查域名健康、攻击、流量、CC/WAF/区域封禁、智能防护或 CCAI 状态时使用。

日本語の概要は準備中です。原文の説明を表示しています。

bytedance/agentkit-samples4712026年10月12日 更新

bytedance のスキルをすべて見る

このスキルの問題を報告する