Introduction

Higgsfield just opened Genjutsu, the motion-transfer tool behind a big slice of its most-shared clips, as a public API. The launch post went up on October 12 with a 50% off promo that runs through October 31, and the Higgsfield API console currently lists Genjutsu from $0.159 per second (down from $0.318). Then, on October 13, ComfyUI merged a Higgsfield Genjutsu Motion Transfer partner node, so you can drop it straight into a graph.

It's closed and API-only, with no weights. You can run it today in the Higgsfield app, through the Higgsfield API, or (once your ComfyUI build includes the new node) as a Comfy partner node.

What Genjutsu does

Genjutsu keeps the motion, camera movement and timing of a real clip and rebuilds everything performing it from your references. Film yourself waving a plush toy, hand it a character sheet and a city street, and the same wave plays out with a new hero in a new place.

The API's motion-transfer endpoint works like this:

  • Source video: at least 4 seconds. Anything past 30 seconds is trimmed, and output length follows the prepared source.
  • References: 1 to 8 images of characters, outfits, props, locations or styles, in order.
  • Prompt: optional, up to 10,000 characters. Use it to say which reference replaces whom, or to describe a new setting.
  • Resolution: 480p, 720p (default) or 1080p.
  • Audio: the ComfyUI node notes that the source clip's audio is kept along with its motion.

Higgsfield's own Genjutsu page also shows Object Swap (change one element, keep the rest of the shot) and batch runs from one setup. The API endpoint covered here is motion transfer.

Call the API

Grab a key ID and secret from the Higgsfield console, host your clip and references at public URLs, then submit:

curl --request POST \
  --url https://api.higgsfield.ai/higgsfield/genjutsu/motion-transfer/v1.0 \
  --header "Authorization: Key ${HF_API_KEY_ID}:${HF_API_KEY_SECRET}" \
  --header "Content-Type: application/json" \
  --data '{
    "video_url": "https://example.com/my-dance-take.mp4",
    "image_urls": [
      "https://example.com/hero-character.jpg",
      "https://example.com/neon-rooftop.jpg"
    ],
    "prompt": "The person in the video becomes the character in image 1, performing on the rooftop from image 2 at night",
    "resolution": "720p"
  }'

The API is asynchronous: you get a request ID back, then poll its status (or use a webhook) until it reports completed and returns a video URL. A few house rules from Higgsfield's docs worth knowing:

  • Failed or moderated requests aren't charged. Reserved credits are refunded automatically.
  • Outputs are kept for at least seven days, so download what you want to keep.
  • Estimate before you burn credits. There's an estimate endpoint that takes the same parameters.
  • Python and TypeScript clients exist (higgsfield_client and @higgsfield/client).

Use it in ComfyUI

The new partner node, Higgsfield Genjutsu Motion Transfer, sits under partner/video/Higgsfield. It landed in ComfyUI's main branch on October 13, after the latest tagged release (v0.39.0), so you'll need a nightly build or the next release to see it.

What the node gives you:

  • Inputs: a video, an auto-growing stack of reference_images (1 to 8), an optional prompt, resolution, and a seed that only controls re-runs (Higgsfield results aren't deterministic).
  • Output: a single VIDEO, ready for Save Video or any downstream video nodes.
  • Billing: it runs through your Comfy.org account like other partner nodes. The node's price badge reads about $0.45/s at 480p, $0.97/s at 720p and $2.33/s at 1080p. That's Comfy's partner-node rate, not Higgsfield's direct promo price, so compare before you queue a long clip.

A fun graph to try: generate a character sheet with your favorite image node, feed it into the Genjutsu node's first reference slot, and point the video input at a short phone clip of yourself. That's a recast short in one queue.

Prompt starter

Genjutsu's prompt is optional, but a clear casting note helps when there are multiple people or references. Paste and edit:

The person in the video becomes the character in image 1, wearing the outfit from image 2. Move the scene to the location in image 3 at golden hour. Keep the original motion, camera move and timing exactly as filmed.

Conclusion

Motion transfer used to mean a heavy local pipeline and a lot of patience. Now it's one endpoint, half price until Halloween, and a single node in ComfyUI. The obvious caution is cost: you pay per second, and it adds up fast at 1080p. Start at 480p, find the take you love, then upscale your ambitions. Grab a 4-second clip, pick your cast, and go make some magic.

Original Source

—Aurelia ♡