SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsvmake-smart-montage

vmake-smart-montage

Design & creation

View repository

Updated 10-16 00:23

About this skill

Smart montage

Capability and preparation

Use mixed-cut analyze / generate to interpret media and the creative goal, arrange visuals and pacing, and generate after plan confirmation. Requires vmake-labs-cli, targeting version 0.1.11, compatible with >=0.1.11 <0.2.0.

Read CLI setup before the first invocation. Check the version and both mixed-cut analyze --help and mixed-cut generate --help. Query the relevant vmake contracts entries when JSON fields or media limits are unclear. For talking-head plan edits, read generation answers; for output, interaction, or errors, read runtime and recovery.

  • Images, videos, and mixed inputs are supported. Multiple images can form a complete video. Follow CLI media validation; the service chooses the appropriate analysis branch.
  • material_mixed_cut uses the actual media arrangement, script, and voice information. talking_mixed_cut uses ARoll video with optional BRoll images/videos and preserves the original speech. ARoll requirements apply only to the talking-head branch.
  • Optional --target-duration specifies a 5-180 second target. Omit it when unspecified. --language supplies content-language context; the talking-head branch retains original speech. --lang sets the client language.
  • Include requested pause handling, pacing, aspect ratio, duration, and other goals in the analysis prompt. Use the real plan and user confirmation. Treat recognized a_roll_text as read-only and deliver the actual returned video.

Collect media and analyze

  1. Collect 1-20 real image/video files or URLs and a theme. Preserve provided copy, ordering, narrative, and pacing. When inputs and theme are clear, prepare and analyze directly; do not require the user to classify ARoll/BRoll first. Ask only for missing information that affects the result.
  2. Use readable local files. Download every remote original according to CLI setup, including session media and historical outputs. Preserve sources, local paths, ordering across types, and repeated inputs. If a download fails, ask for the original local file; analyze only once all selected inputs are available. For more than 20 inputs, help the user select without silently truncating, batching, or choosing.
  3. Arrange --image-file and --video-file in the user's order, repeating flags as needed. Do not regroup by media type. Add -r '<session_id>' to reuse a session; omit it for a new one. Put the theme and known requirements in -p; use current help for optional flags.
  4. The CLI validates the whole batch, uploads in order, and submits one analysis request using its CDN URLs. Do not first call create-room or add-media, or use --item-id. Describe roles by positions such as "the first input"; local paths are not remote references. Do not invent unseen scenes or edit points.

This example mixes an image and a video. Adjust file arguments to the actual inputs without changing their order; image-only and video-only input are also supported:

vmake mixed-cut analyze --image-file '<local-image-path>' --video-file '<local-video-path>' -p '<user-theme-and-requirements>'

Analysis creates a plan, not the final video. Preserve actual returned session and Task IDs and the original media mapping. Follow CLI setup for defaults, login, and argument handling. Parameter files and separate validation are optional.

Wait for and confirm the plan

  1. Read data.type=mixed_cut_brief from actual Message Parts or Task status.message Parts. Update by (session_id, brief_id); do not guess IDs from prose, numbering, or examples.
  2. While pending, query history or known Tasks read-only. Even if history exits as idle, do not generate. Space serial queries; when no update arrives for a prolonged period, explain the state and preserve recovery details without repeatedly analyzing. Stop on failed. If a Brief is absent, recover history first and report the problem if it remains missing. A command ending, a completed Run, or Session idle does not mean the Brief is ready.
  3. At ready, show the plan for the actual brief_type. For talking-head, show the main video, supporting media, target duration, and read-only a_roll_text. For material montage, show actual media arrangement, script, voice_id, and target duration. Do not invent absent fields or claim to have seen unavailable footage.
  4. Images returning a material plan are normal. For an unknown branch or missing required information, follow the actual question without changing the type or fabricating fields.
  5. Obtain explicit confirmation after displaying the ready plan. A general request such as "edit this" cannot approve an unseen plan. When editing the plan, explain the final changes; do not ask again for the same plan already explicitly accepted.

Generate and edit

Use the confirmed real session and ready Brief. Both branches use the same command; omit --answer for the material branch, which does not require ARoll:

vmake mixed-cut generate -r '<session_id>' --brief-id '<brief_id>'

Submit after valid login and user confirmation of the current ready Brief. Local validation is optional for diagnosis; it cannot establish that the Brief exists, is ready, belongs to the inputs, or has been confirmed. The CLI submits generation directly without querying history first, so the caller must check the actual analysis result.

  • Only the talking-head branch accepts --answer edits to a_roll_media, b_roll_media, and target_duration. Media items contain real short_id and kind only. ARoll requires at least one video; an empty BRoll array explicitly clears supporting footage. Omit unchanged fields. Serialize a direct-flag answer as one argument, or use an answer object in a parameter file. See generation answers.
  • Choose only inputs already analyzed for the current plan. Do not pass type, a_roll_text, script, voice_id, confirmed_fields, URLs, or local paths. Submit the answer with the generation action.
  • For new or replacement source media, or changes to a material plan's theme, copy, selections, or duration, download remote originals first and reanalyze the full local input list with updated requirements in the same session. Present and confirm the new Brief before generating. Do not preregister new media.
  • reply --approve handles the current permission request; reply --answer answers an actual native question. Neither replaces Brief generation or these plan edits.

Follow and deliver

  • Read the original stdout and stderr separately, retain all tasks, and avoid a second concurrent observer. Deliver only with TASK_STATE_COMPLETED and a usable current video URL. HTTP 200, exit code 0, end.reason=result, and unknown states do not alone prove success.
  • Deliver every actual usable video promptly, keeping complete URLs and viva.short_id. Include an actual room_url if returned. Use vmake-dynamic-caption only if captions are requested and the Skill is installed; otherwise explain the equivalent caption workflow without silently adding processing.
  • Query Task/history for missing or expired URLs; a media ID or cover is not the video. Recover read-only after disconnection, timeout, or uncertain acceptance. Retry analysis or generation only after fixing a confirmed failure and obtaining authorization.
  • A caption or download failure must not repeat a successful montage. Deliver successful results and retain recovery context. Remove task-created temporary media or parameter files once no longer needed, preserving user originals. See runtime and recovery for saving and recovery details.