Seedance Volcengine Workbench

Local web app for Seedance 2.0 on Volcengine Ark. It includes:

• Ark API key checker
• server-side in-memory key pool
• round-robin key rotation
• Seedance 2.0 / Seedance 2.0 Fast task creation
• task polling and video preview
• optional Qwen Console for HappyHorse 1.1 text-to-video and image-to-video
• clipboard image/video paste for reference media inputs
• automatic Litterbox upload for pasted reference videos
• optional autosave of generated videos to a local folder
• up to 9 Seedance 2.0 reference images

Setup

npm install
cp .env.example .env
npm run dev

Open the Vite URL shown in the terminal. In development, the frontend runs on port 5173 and proxies /api to the Express backend on port 8790.

You can provide keys in .env with ARK_API_KEYS, or paste them into the Key Pool panel and activate them from the UI. Keys are stored in backend memory only and are never returned unmasked by the API.

Qwen Console is disabled by default. To show it, set QWEN_CONSOLE_ENABLED=true and DASHSCOPE_API_KEY in .env. It can submit HappyHorse 1.1 text-to-video tasks with happyhorse-1.1-t2v and first-frame image-to-video tasks with happyhorse-1.1-i2v, then poll task IDs from the same panel. Override DASHSCOPE_VIDEO_ENDPOINT or DASHSCOPE_TASK_ENDPOINT only if your DashScope account uses a different region endpoint.

Backend outbound requests go direct by default. To enable a proxy for Ark calls, public media URL preflight, Litterbox uploads, and autosave downloads, set APP_PROXY_URL or ARK_PROXY_URL to a value such as socks5://127.0.0.1:2080; APP_PROXY_URL wins when both are set. Set the chosen value to off to force direct connections.

For first-frame image-to-video, focus the Generation panel and paste an image with Ctrl+V, or use the Paste button in the Image field. Public URLs still work in the same field.

For Seedance 2.0 reference images, use the Reference Images field. It supports pasted images, public image URLs, and asset:// IDs. Each entry is sent as an image_url content item with role: "reference_image", up to 9 images.

Pasted images can be submitted directly. Pasted videos are previewed immediately, uploaded to Litterbox, and then submitted as the returned public video URL. The default Litterbox expiry is 12h; set LITTERBOX_EXPIRY to 1h, 12h, 24h, or 72h to change it.

The backend performs a small byte-range preflight for public Video and Audio URLs before creating an Ark task. If Ark still reports timeout while fetching resource, the file host is not reachable fast enough from Volcengine Ark; rehost the media on a stable public bucket/CDN or use an asset:// ID.

Video and Audio reference media are sent with reference_video and reference_audio. Manual Video URL entries can still be public video URLs or asset:// IDs.

Use the Autosave panel to enable local saving of generated videos. Choose a folder with the system dialog or enter a writable folder path manually; settings save automatically. Completed task videos are downloaded by the backend after polling returns content.video_url. The setting is stored in .autosave-settings.json, and .env can provide initial defaults with AUTOSAVE_ENABLED and AUTOSAVE_DIR.

Ark Defaults

• Base URL: https://ark.cn-beijing.volces.com/api/v3
• Create task: POST /contents/generations/tasks
• Query task: GET /contents/generations/tasks/{id}
• Key check: GET /contents/generations/tasks?page_num=1&page_size=1

Qwen Defaults

• Submit task: DASHSCOPE_VIDEO_ENDPOINT
• Query task: GET {DASHSCOPE_TASK_ENDPOINT}/{task_id}
• Models: happyhorse-1.1-t2v, happyhorse-1.1-i2v
• Resolutions: 720P, 1080P

Model Watchtower

The watchtower polls Ark /models endpoints, tracks models matching WATCHTOWER_MODEL_FILTER, and sends Telegram alerts when matching models are added, removed, or changed.

Add these to .env:

TELEGRAM_BOT_TOKEN=123456:telegram-token
TELEGRAM_CHAT_ID=123456789
TELEGRAM_COMMANDS=true
ARK_PROXY_URL=socks5://127.0.0.1:2080
WATCHTOWER_MODEL_FILTER=seed
WATCHTOWER_INTERVAL_MS=60000
WATCHTOWER_CONCURRENCY=64

To check every key in ../keychecker/working_doubao_models.txt, use the key-file scripts:

npm run watchtower:keys:dry-run
npm run watchtower:keys:once
npm run watchtower:keys

Key-file mode queries every key independently with hashed key labels in the snapshot and alerts. The bundled key-file scripts set ARK_PROXY_URL=socks5://127.0.0.1:2080 and WATCHTOWER_IGNORE_ENV_KEYS=true, so only keys from that file are checked through the local SOCKS5 proxy. Its default state file is .watchtower/ark-seed-models-all-keys.json.

All-key alerts are grouped by model/change shape by default, so a new model visible on many keys is sent as one line with keys=<count> and sample hashed key labels instead of one line per key. Use WATCHTOWER_AGGREGATE_CHANGES=false only when you want raw per-key detail. WATCHTOWER_MAX_CHANGE_LINES=120 caps the detailed change lines in one alert.

The all-key scan uses WATCHTOWER_CONCURRENCY=64 by default. On the local SOCKS5 proxy this cuts the 56-key, two-endpoint check from about 20 seconds to about 5 seconds.

Telegram bot commands are registered automatically when TELEGRAM_BOT_TOKEN and TELEGRAM_CHAT_ID are set and the continuous watcher is running:

/status - show watcher status
/check - run a model check now
/models - list tracked Seed model groups
/errors - show current key/API errors
/reset confirm - delete the current baseline
/ping - check bot responsiveness
/help - show commands

Only TELEGRAM_CHAT_ID is allowed to execute commands. Command polling stores its offset in .watchtower/telegram-offset.json so old commands are not replayed after restart.

Useful commands:

# Seed-family dry run without writing state or sending Telegram
npm run watchtower:dry-run

# One check using the normal Ark key pool
npm run watchtower:once

# Continuous watcher using the normal Ark key pool
npm run watchtower

# Dry run across every key in ../keychecker/working_doubao_models.txt
npm run watchtower:keys:dry-run

# One check across every key and write/update the all-key baseline
npm run watchtower:keys:once

# Continuous all-key watcher, through socks5://127.0.0.1:2080
npm run watchtower:keys

# Reset the normal baseline; next run creates a fresh baseline
npm run watchtower:reset

# Reset the all-key baseline; next all-key run creates a fresh baseline
npm run watchtower:keys:reset

The first non-dry run writes .watchtower/ark-seed-models.json as the baseline and does not alert unless WATCHTOWER_NOTIFY_INITIAL=true. Later runs compare against that snapshot and send Telegram messages for changes.

PowerShell helpers:

# Run one all-key check with a Telegram token/chat id only for this process
$env:TELEGRAM_BOT_TOKEN="123456:telegram-token"; $env:TELEGRAM_CHAT_ID="123456789"; npm run watchtower:keys:once; Remove-Item Env:TELEGRAM_BOT_TOKEN; Remove-Item Env:TELEGRAM_CHAT_ID

# Ask Telegram for recent chat ids through the local SOCKS5 proxy
curl.exe -sS --socks5 127.0.0.1:2080 "https://api.telegram.org/bot$env:TELEGRAM_BOT_TOKEN/getUpdates"

Notes

Seedance task URLs are provider-hosted assets. Save finished videos promptly if you need durable storage.