Vidjutsu Clone

vidjutsu-clone

Recreate a TikTok video's motion on a new character end to end, driving only the VidJutsu API. Every step in this skill — download, cloneability check, character generation, starting-image generation, motion clone, and status poll — is a deterministic VidJutsu endpoint. VidJutsu holds the WaveSpeed and Gemini credentials on its own side; you never see or hold them.

When to invoke

• The user has a TikTok URL and wants that video's motion recreated on a different character (a new face/persona doing the same moves, same sound)
• They want the full chain run for them — download through a finished, hosted video URL — without touching a provider SDK directly
• The source clip is 15 seconds or shorter (VidJutsu caps clone sources at 15s)

This is not the same job as clone-ad (sibling skill, same repo). clone-ad clones an ad's pacing and shot structure for a different product using Gemini + Wavespeed directly, producing a fresh multi-shot 15s spot. vidjutsu-clone clones a single performer's motion from one source clip onto one character image, through VidJutsu's motion-clone endpoints only. If the user wants a multi-shot product ad, use clone-ad. If they want "make this TikTok but with a different person doing it," use this skill.

Required environment variables

• VIDJUTSU_API_KEY — required. This is the only key this skill uses. Set it once via vidjutsu auth (writes ~/.vidjutsu/config.json) or export it as VIDJUTSU_API_KEY in the shell. Do not ask the user for a WaveSpeed key, a Gemini/GOOGLE_ key, or any other provider credential — VidJutsu's clone endpoints call those providers server-side under VidJutsu's own request-scoped credentials and never expose them to the caller.

Stop before running anything if VIDJUTSU_API_KEY isn't set. Point the user to curl -fsSL https://vidjutsu.ai/install.sh | bash followed by vidjutsu subscribe --email you@example.com (see below) rather than attempting a partial run.

Output

Save the download source, character image, starting image, and final clone video URL as a JSON manifest under CWD/vidjutsu-clone-<timestamp>/result.json so a run is never overwritten and outputs are easy to find and clean up.

Primary path: the CLI (vidjutsu clone run)

Prefer this. It is the same deterministic chain described below, wrapped in one command:

vidjutsu clone run "https://www.tiktok.com/@handle/video/1234567890" \
  --character-prompt "A neutral, camera-ready presenter" \
  --model kling

What it does, in order:

1. POST /v1/videos/download/tiktok — pulls the TikTok video onto VidJutsu's CDN and returns a stable url + assetId.
2. POST /v1/clones/check — scores cloneability. Prints the verdict, a 0–100 score, and evidence. If the verdict is weak, the command stops before spending anything on generation and exits nonzero. Pass --force only if the user explicitly wants to proceed anyway, and say so out loud — a weak verdict means the clone is unlikely to look good.
3. POST /v1/characters — generates a reusable character identity image from --character-prompt (default: "A neutral, camera-ready presenter" if the user hasn't described one).
4. POST /v1/clones/starting-image — composes a clean, text-free opening frame that swaps the character into the source video's framing and pose.
5. POST /v1/clones/video — submits the motion clone. Kling motion control is the default model (puppets the character image with the source video's motion, keeps the original sound). Pass --model seedance for the alternate model (480p, 9:16, VidJutsu can rewrite the prompt).
6. Polls GET /v1/clones/video/{id} until status is completed (or failed), then prints the final videoUrl.

Install if the CLI isn't present yet:

curl -fsSL https://vidjutsu.ai/install.sh | bash
vidjutsu subscribe --email you@example.com   # only if not already subscribed

The CLI resolves the key from VIDJUTSU_API_KEY, then ~/.vidjutsu/config.json — nothing else to configure.

Fallback path: raw curl chain

Use this only if the CLI isn't installed and installing it isn't an option. Every call needs Authorization: Bearer $VIDJUTSU_API_KEY.

# 1. Download the TikTok source onto VidJutsu's CDN
download=$(curl -fsS -X POST https://api.vidjutsu.ai/v1/videos/download/tiktok \
  -H "Authorization: Bearer $VIDJUTSU_API_KEY" -H "Content-Type: application/json" \
  -d "{\"url\":\"$TIKTOK_URL\"}")
source_url=$(echo "$download" | jq -r '.url')

# 2. Cloneability check — branch on the verdict before spending on generation
check=$(curl -fsS -X POST https://api.vidjutsu.ai/v1/clones/check \
  -H "Authorization: Bearer $VIDJUTSU_API_KEY" -H "Content-Type: application/json" \
  -d "{\"videoUrl\":\"$source_url\"}")
verdict=$(echo "$check" | jq -r '.verdict')
if [ "$verdict" = "weak" ]; then
  echo "Cloneability verdict is weak — stopping. Evidence:"
  echo "$check" | jq -r '.evidence[]'
  exit 1
fi

# 3. Generate the character
character=$(curl -fsS -X POST https://api.vidjutsu.ai/v1/characters \
  -H "Authorization: Bearer $VIDJUTSU_API_KEY" -H "Content-Type: application/json" \
  -d "{\"prompt\":\"A neutral, camera-ready presenter\"}")
character_url=$(echo "$character" | jq -r '.imageUrl')

# 4. Generate the starting image (character swapped into the source framing)
starting=$(curl -fsS -X POST https://api.vidjutsu.ai/v1/clones/starting-image \
  -H "Authorization: Bearer $VIDJUTSU_API_KEY" -H "Content-Type: application/json" \
  -d "{\"characterImageUrl\":\"$character_url\",\"prompt\":\"Match the framing and pose of the source video's opening frame\",\"sourceVideoUrl\":\"$source_url\"}")
starting_url=$(echo "$starting" | jq -r '.imageUrl')

# 5. Submit the motion clone (kling is the default model; pass "model":"seedance" for the alternate)
submit=$(curl -fsS -X POST https://api.vidjutsu.ai/v1/clones/video \
  -H "Authorization: Bearer $VIDJUTSU_API_KEY" -H "Content-Type: application/json" \
  -d "{\"startingImageUrl\":\"$starting_url\",\"sourceVideoUrl\":\"$source_url\",\"model\":\"kling\"}")
task_id=$(echo "$submit" | jq -r '.id')

# 6. Poll to completion
while true; do
  status=$(curl -fsS https://api.vidjutsu.ai/v1/clones/video/$task_id \
    -H "Authorization: Bearer $VIDJUTSU_API_KEY")
  state=$(echo "$status" | jq -r '.status')
  [ "$state" = "completed" ] && { echo "$status" | jq -r '.videoUrl'; break; }
  [ "$state" = "failed" ] && { echo "$status" | jq -r '.error'; exit 1; }
  sleep 3
done

Do not poll in a tight loop with no delay — 3s between checks is enough and avoids hammering the endpoint.

Model choice

• Kling motion control (default) — puppets the character with the source's motion and keeps the original audio untouched. Preserve its output exactly as returned; don't resize, transcode, or otherwise normalize the pixels.
• Seedance (alternate, --model seedance) — video-edit model, submitted at 480p / 9:16. Accepts an optional prompt override for the identity swap if the default composition instruction isn't specific enough.

Don't switch models mid-run once one has been picked — regenerating with a different model after a failure should be a deliberate retry, not a silent fallback.

Constraints

• Source clips are capped at 15 seconds. Longer TikToks will fail the download or check step — tell the user to trim first if their source is longer.
• Stop on a weak cloneability verdict. Don't spend on character, starting-image, or video generation past a weak verdict unless the user has explicitly said to proceed anyway.
• One motion-clone attempt at a time per source. If a clone comes back wrong, retry the same model at most twice before telling the user it isn't working and asking how they want to proceed — don't loop indefinitely.

Cost

Every call in this chain — download, check, character, starting-image, video, and the status poll — is metered against your VidJutsu account, not billed separately per provider. VidJutsu is a flat monthly subscription with per-endpoint daily rate limits (reset 00:00 UTC), not a pay-per-call credit system. GET /v1/pricing (or vidjutsu pricing if the CLI exposes it) is the current source of truth for the subscription price and daily limits per endpoint — check it before running this at volume so you don't hit a 429 mid-chain. Status polls (GET /v1/clones/video/{id}) are not billed.

The chain takes about 5 minutes end to end once submitted, mostly waiting on the motion-clone step.

No provider keys, ever

This skill talks to api.vidjutsu.ai only, authenticated with VIDJUTSU_API_KEY. VidJutsu's clone endpoints call WaveSpeed (character and starting-image generation, Kling/Seedance motion clone) and other model providers on its own side, using its own request-scoped credentials. Never ask the user to set WAVESPEED_API_KEY, GEMINI_API_KEY, a GOOGLE_* variable, or any other provider key for this skill — if a step ever appears to need one, that's a sign the wrong endpoint is being called.