vidu-skills

Vidu Generation and Asset Workflows

Use vidu-cli v0.2.36 or newer. This skill is aligned with v0.2.36 (6bb4b07). Check vidu-cli --version on first use; ask users of older versions to upgrade rather than inventing compatibility fallbacks.

Execution Model

Call the CLI with the host's command execution tool. Prefer argv arrays when supported; if the host accepts shell commands, quote arguments correctly and use prompt/timeline files for complex content. No batch orchestration tool is required. Use CLI flags, not raw API request bodies; compose timeline JSON is an input supported by the CLI.

Keep the default JSON output for agents. --format table is for human inspection. Parse exit status, stdout, and stderr separately: argument errors are text on stderr; runtime failures normally return ok: false JSON. See errors and retry.

Environment: VIDU_TOKEN is required. VIDU_BASE_URL defaults to https://service.vidu.cn; overseas accounts use https://service.vidu.com. Optional VIDU_DEBUG=1 prints response bodies to stderr. Read setup for installation or credential problems.

Choose a Workflow

User intent	Command / reference
Generate images or videos	task submit --type ... --model-version ...; read parameters for the task/model matrix, reference limits, and defaults.
Generate speech or lip-sync	task tts or task lip-sync; read parameters. Select official voice IDs from task tts-voices / task lip-sync-voices.
Compose video/audio/subtitle/effect tracks	task compose --timeline ...; read compose before building a timeline.
Find, inspect, wait for, or download a task	task list, task get --details, task wait; read task workflows.
Cancel/delete a task, rerun, extend, or enhance a creation	task cancel/delete/rerun/extend/upscale-image/upscale; read task workflows.
Create/manage reference elements or browse/favorite creations	element ..., creation ...; read assets.
Clone/manage a personal voice and use it for TTS	voice create/list/get/update/wait; read voices.
Check quota, credits, transactions, or a quote	quota pass/credit/records, task cost/tts-cost/lip-sync-cost; read parameters. Video upscale quotes use task upscale-cost, documented in task workflows.
Upload a reusable image	upload <local_image_path> returns upload_id and ssupload_uri; image input flags can also upload automatically.

Core Rules

• Respect the user's selected model. Do not switch defaults solely because a newer model is available. Read the model matrix before adding model-specific settings.
• Repeat --image, --material, --audio, and --video per input. Each material is one name:id:version argument; never comma-join materials.
• Input requirements are model-specific. Q4 supports audio-only references; Drama and Wan 3 accept multimodal references with different limits. Do not apply the legacy image/material count to every reference task.
• --audio supplies reference audio; --with-audio enables output audio and --with-audio=false disables it on supported video models. Omission preserves existing behavior. Quote and submit with the same audio selection.
• Schedule modes: nopeak = off-peak mode (闲时模式), claw_pass = pass mode (通行证模式), normal = credit mode (积分模式). Availability and off-peak eligibility/quota depend on the model and server. Do not automatically change an explicitly selected mode after rejection.
• Image tasks need no duration; element create has no duration option. Drama/Wan 3 have video defaults; older submit models still require explicit duration/resolution as applicable.
• --prompt-file aliases --prompt-path; --model aliases --model-version; --voice aliases --voice-id where supported. Prefer full names in examples.
• Task IDs, numeric creation IDs, element IDs/versions, and personal voice_... IDs come from responses and are not interchangeable.
• Generation returns a task_id by default. When the user needs a finished artifact, prefer --wait --output ./results for a single task. For a batch, submit each task, save all IDs, then wait individually.
• On task submissions, --timeout and --output require --wait. Use a directory for multiple artifacts or audio plus subtitles; a file path is allowed for a single artifact.
• Waiting timeout/interruption does not cancel the server task. Continue with the original ID. Never blindly resubmit after an ambiguous submission failure, or change a failed voice trial into a paid clone.

Common Shapes

# The user's chosen model; legacy required settings remain explicit.
vidu-cli task submit --type text2video --prompt "A cat walking through rain" \
  --model-version 3.2 --duration 5 --resolution 1080p --wait --output ./results

# Explicit request for a Wan 3 video with generated audio.
vidu-cli task submit --type text2video --prompt "Waves breaking on a beach" \
  --model-version viduq2-drama-std --with-audio --wait --output ./results

# Resume an existing task; TASK_ID must come from a response.
vidu-cli task wait "$TASK_ID" --timeout 1800 --output ./results

There is no task download, top-level reference, voice delete, or task submit --type tts command. Use only the supported workflows above. If the installed CLI differs, inspect the relevant --help; do not guess flags or silently change the user's model or billing choice.