Higgsfield Image-to-Video (Direct API)

Image-to-video is the reason most people stumble onto this pack, and this is the node they were looking for. Give it a still, describe the motion, and it hands you back a finished clip from Kling 2.1 Pro, ByteDance's Seedance Pro, or Higgsfield's own DOP - three of the closed models the local community measures its Wan and LTX rigs against, all without you renting a 24GB card or waiting out a ten-minute local render.

The catch is the same one as the pack's other nodes: it's a paid hosted API. Every successful clip costs real money. But for a quick client deliverable or a "what if I animated this?" experiment, a few seconds of Kling on someone else's hardware is frequently the pragmatic answer - the quality floor is high and the wait is usually shorter than a local queue.

How it works

The node takes your IMAGE, uploads it to Higgsfield (as a JPEG), then calls higgsfield_client.subscribe() with the model, your motion prompt, and the duration. Once the job finishes it downloads the resulting mp4 straight into your ComfyUI output folder - this matters more than it sounds. Higgsfield only keeps generated files for a minimum of seven days, after which the URL dies. The node deliberately pulls the bytes down to disk so your result outlives the link, which is the difference between "I have a video" and "I have a link that expires next week."

The inputs

• image - your starting still. The first frame you feed it is basically the composition you'll get, so frame your shot before you press go.
• prompt - describe the motion, not the scene. "Slow push-in, rain starting, hair moving" beats "a woman standing in the street." You're paying per attempt, so a good motion prompt is literally money saved.
• model - defaults to higgsfield-ai/dop/preview. The "preview" in the name is a hint that it's the cheaper/experimental lane; bytedance/seedance/v1/pro/image-to-video (Seedance Pro) and kling-video/v2.1/pro/image-to-video (Kling 2.1 Pro) are the heavyweights. If you want the flashiest result, reach for Kling; if you want to spend less while testing, DOP is a fine starting point.
• duration - 5, 10, or 15 seconds. Longer clips cost more and give the model more room to drift. Start at 5.

Optional api_key and auto_save round it out; flip auto_save off and the file lands in the temp folder instead, cleared with the session.

Outputs

Four slots: video_url (the remote link - treat it as short-lived, the 7-day expiry applies), cache_key (the deterministic fingerprint for ComfyUI-API-Optimizer's Hash Vault), info (status string with model, duration, timing, saved path), and - the good one - video, a native ComfyUI VIDEO you can wire straight into any video-compatible node. The node is marked as an output node, so it'll also show its progress in the UI.

Install and gotchas

Install is the pack's shared routine: ComfyUI Manager (search "Higgsfield") or

cd ComfyUI/custom_nodes
git clone https://github.com/jeremieLouvaert/ComfyUI-Higgsfield-Direct.git
pip install -r ComfyUI-Higgsfield-Direct/requirements.txt

then add your key via the higgsfield_api_key.txt file in your ComfyUI root, an HF_KEY environment variable, or the node's api_key input. Key comes from cloud.higgsfield.ai. Everything lives under AKURATE/Higgsfield.

Where people get burned: the VIDEO output type needs a recentish ComfyUI (the comfy_api module arrived mid-2025). On older builds the slot comes back empty - but the node still saves the file to disk, so you're not out a generation. The real cost trap is retries: every regen is a new charge, and video is where "let me try one more seed" gets expensive fast. And don't forget the content moderation - it's a hosted consumer service, so anything spicy is getting rejected, which is at least not charged.