video-batch-runner（视频批量运行器）

Video Batch Runner

Follow shared release-shell rules in:

• postplus-shared release-shell rules

Use this skill after image, script, voice, or prompt-planning work already exists.

This skill is for:

• generating talking-head videos from approved image and audio inputs
• generating Seedance videos from text, images, videos, audio references, or first/last frames
• storing render jobs as local assets with normalized manifests
• preserving traceability from final video back to persona, concept, script, and voice take
• keeping render providers replaceable behind a stable adapter contract

This skill is not for unconstrained video ideation.

Quality Default

When the goal is believable human video, default to the highest practical render quality the provider offers.

Default quality assumption:

• lower render quality can make already-imperfect source faces look more fake
• realism-sensitive talking-head jobs should start from the best available resolution before blaming script or voice
• only step down when the user is explicitly running a cheap draft, a latency test, or a provider-limited experiment

The selected resolution should always be persisted in the request and manifest.

Core Idea

The video layer should be organized around render objects, not around one provider endpoint.

Treat a video render as:

• render job

- one attempt to produce a video from approved upstream assets

• render manifest

- the normalized local record of that attempt

• video asset

- the downloaded output file(s) produced by the render

The main value is not "image plus audio becomes video". The main value is preserving the chain:

• campaign
• concept
• persona
• script
• voice take
• image asset
• render job
• qa report

Fact Rule

Video render inputs must be grounded in approved upstream assets or an explicit prompt plan.

Required upstream inputs depend on route:

• talking-head

- approved image asset - approved voice take - script or concept reference - render purpose - local output directory

• ark / seedance

- prompt or promptPlan - any required media refs for the chosen mode - concept reference - render purpose - local output directory

Do not let the render stage silently redefine:

• who the persona is
• how the voice should sound
• what the concept is trying to test

If the request is experimenting with a new render style, record that as an explicit render variant.

Source Selection Rule

Start from the active project's approved upstream assets and manifests.

If a current task clearly belongs to one project or client folder, stay within that context first.

Do not assume one client directory is the default home for all renders.

Video Routes

Current routes:

• talking-head

- model: hosted talking-head capability - category: image-to-video digital human

• seedance (hosted)

- endpoint keys: video-seedance-2-image, video-seedance-2-image-turbo, video-seedance-2-text, video-seedance-2-text-turbo - category: text/image/reference-media to video

• ark

- direct workspace route for internal video/audio workflows - category: text/image/video/audio to video

Read references/hosted-video-talking-head.md before implementation or request design. Read references/hosted-video-generative.md before designing hosted Seedance requests. Read references/volcengine-seedance-2.md before designing Seedance requests.

If the project should keep related image, audio, and video files under one asset root, use the shared asset model in ../image-batch-runner/references/unified-asset-contract-v1.md.

Hosted Boundary Rule

• keep request files, raw provider responses, and polling state under <work-folder>/.postplus/video-batch-runner/ when they are internal execution state
• keep only final user-facing renders outside .postplus/
• if hosted video capability is unavailable, unauthorized, or returns a stable network error, stop immediately instead of switching to ad hoc shell glue

Render Objects

1. Render Job

One request to a video provider.

Should include:

• jobId
• campaignId
• personaId
• conceptId
• scriptId or source path
• voiceTakeId
• imageAssetId
• assetPurpose
• provider
• model
• status

2. Render Manifest

The normalized local handoff object for later review.

Should include:

• jobId
• campaignId
• personaId
• conceptId
• provider
• model
• requestPath
• responsePath
• predictionId
• providerStatus
• assets[]
• sourceBasis
• upstreamRefs

3. Video Asset

One downloaded output.

Should include:

• assetId
• localPath
• remoteUrl
• mimeType
• sourceBasis
• createdAt

Default Workflow

1. Lock the render brief

Before calling any provider, write down:

• jobId
• campaignId
• personaId
• conceptId
• scriptId or script source
• voiceTakeId
• imageAssetId
• assetPurpose

- talking_head - singing_avatar - first_pass_render - render_fix

• sourceBasis
• mustKeep
• canVary
• feedback

2. Produce a normalized request record

The local request JSON should contain stable fields even if provider fields change later.

At minimum record:

• provider route
• model
• prompt or prompt plan
• media refs used by the route
• optional mask image
• resolution
• ratio when relevant
• duration or frames when relevant
• seed
• local output directory

When the provider exposes multiple resolution tiers, default to the highest practical tier for realism-sensitive renders.

3. Call the provider and save raw response

Always save:

• request.json
• response.json
• manifest.json
• downloaded video files under renders/

Do not use the provider response alone as the durable store.

4. Normalize local outputs

Every run should end with a local manifest containing:

• stable upstream refs
• provider ids
• local asset paths
• source basis
• feedback history

5. Hand off to human QA

Do not auto-approve a render.

The next stage is creative-qa, where a person may record:

• verdict
• what worked
• what failed
• which stage should be rerun

If there is no human feedback yet, the render can remain in review_pending.

Path Selection Rule

Write outputs into the active project's render structure when one already exists.

If no project structure exists yet, choose a clear workspace output path and make it visible in the task summary.

If the chosen location will become the long-term handoff point for a client, prefer confirming the destination with the user.

Example Persistence Convention

One possible project-local layout is:

videos/<job-id>/
  request.json
  response.json
  manifest.json
  renders/
  qa/

Do not assume this example layout is the universal default.

Keep draft request files, raw provider responses, and polling state under <work-folder>/.postplus/video-batch-runner/ when they are internal execution artifacts rather than the final handoff.

Tool Contract

This skill expects these adapters:

• generate_video_from_image_audio
• poll_prediction for async providers

For Seedance, the same generate_video_from_image_audio script is used as the normalized submit entrypoint even though the request may be text-to-video or multimodal; the route is chosen by provider.

The normalized request and manifest shapes live in references/tool-contracts.md.

Review Rule

Before calling a video provider, verify:

• for talking-head, persona is approved
• for talking-head, image asset is approved
• for talking-head, voice take is approved
• for seedance (hosted), request mode is explicit and required media exists
• for ark, request mode is explicit
• for ark, media roles match the intended Seedance mode
• for ark, prompt or prompt plan is concrete enough to constrain the generation
• render request is tied to a real concept or asset purpose
• local output path is explicit

After generation, review:

• lip sync acceptability
• persona continuity
• audio and image match
• TikTok-native feel
• ad-like drift

Failure Mode

Stop and say the request is under-specified if any of these are missing:

• for talking-head, no approved image asset
• for talking-head, no approved voice take
• for seedance (hosted), no prompt or no required first image
• for ark, no prompt or no usable content
• no asset purpose
• no source basis
• no local output path

Do not compensate for missing upstream approvals by letting the render model improvise.

Seedance Prompt Rule

For Seedance 2.0 work, prefer a structured prompt plan over a single dense paragraph.

The adapter accepts promptPlan and turns it into one compact prompt string in this order:

• subject + action
• scene / environment
• camera / shot / motion
• visual style / realism target
• sound intent
• continuity constraints
• must-keep
• must-avoid
• reference bindings such as [图1]...，[图2]...

This is a better default than freehand adjective stacks.

Core Scripts

• scripts/generate_video_from_image_audio.mjs
• scripts/poll_prediction.mjs

These scripts take normalized request JSON files and write:

• request.json
• response.json
• manifest.json
• downloaded videos under renders/