Higgsfield video generation in Nebula — text-to-video and image-to-video short cinematic clips via the DoP ("Depth of Presence") model plus Kling v2.1 Pro and Seedance v1 Pro guest models, with duration and aspect-ratio control. Activate when the user configures the `higgsfield` node or asks about Higgsfield in Nebula. The node is video-output only (no Soul images, Speak lipsync, or motion-camera presets are wired). Sourced 2026-06-04 from the official Higgsfield API docs (docs.higgsfield.ai) plus the Nebula audit guide docs/api-guides/higgsfield.md.
Higgsfield video generation in Nebula — text-to-video and image-to-video short cinematic clips via the DoP ("Depth of Presence") model plus Kling v2.1 Pro and Seedance v1 Pro guest models, with duration and aspect-ratio control. Activate when the user configures the `higgsfield` node or asks about Higgsfield in Nebula. The node is video-output only (no Soul images, Speak lipsync, or motion-camera presets are wired). Sourced 2026-06-04 from the official Higgsfield API docs (docs.higgsfield.ai) plus the Nebula audit guide docs/api-guides/higgsfield.md.
Higgsfield Skill
When to use
User drags or configures the higgsfield node on the canvas.
User wants a short cinematic video from a text prompt (text-to-video).
User wants to animate a still image into motion (image-to-video).
User asks which Higgsfield model to pick — DoP Standard vs. DoP Preview vs. Kling vs. Seedance.
User asks about Higgsfield auth, the key:secret pairing, or why an image input isn't being used.
User asks for Higgsfield features that Nebula does NOT expose (Soul text-to-image, Speak lipsync, motion/camera presets, character references) — answer from the Capability boundaries section so you don't over-promise.
Universal rules
Auth header.Authorization: Key {HIGGSFIELD_API_KEY}. Env var is HIGGSFIELD_API_KEY in the backend .env. Restart the backend after adding it. Higgsfield issues credentials as a key + secret pair; the API expects Key {key}:{secret}. The handler does not split or recombine anything — it passes whatever you store verbatim after the literal Key . So if you have both, store the combined value HIGGSFIELD_API_KEY=key:secret; if you only have a single token, store it alone.
Base URL.https://platform.higgsfield.ai (this is the API host used by the handler and node apiEndpoint). Note the Higgsfield dashboard where you mint a key is cloud.higgsfield.ai — different host; don't confuse them.
Execution pattern: async-poll (confirmed in backend/handlers/higgsfield.py). The node submits the job, then polls every 3 seconds, up to 300 polls (~15 min wall-clock ceiling) before timing out. There is no streaming and no webhook — Nebula polls. Expect a minute or more for longer/higher-fidelity renders; this is normal, not a hang.
Submit: POST {base}/{model_id} with a JSON body → response {"request_id": "...", "status": "queued", ...}.
Poll: GET {base}/requests/{request_id}/status → {"status": "...", ...}.
On completed: response carries video.url; the handler downloads the MP4 into the run dir and returns it on the video port.
Status / error handling (from the poll loop):
queued, in_progress → keep polling.
completed → success; pull video.url. If completed but no URL, the handler raises.
nsfw → raises "content policy violation". Reword the prompt.
cancelled → raises "generation was cancelled".
Submit HTTP codes: 200/201/202 accepted; anything else raises Higgsfield submit failed. A 401/403 here means the key (or key:secret) is wrong or missing.
Input-URI rule (the big gotcha). The optional image input is forwarded to the API as image_urlonly if its value starts with http:// or https://. A bare local file path or data URI is silently dropped — the job runs as pure text-to-video with no warning. Always wire a node that emits a hosted image URL into the image port. The handler does not auto-upload local files (unlike the Meshy/Runway handlers, which convert local paths to data URIs).
Key gotchas.
prompt is required for every model, including the image-to-video ones — Kling/Seedance still need a motion prompt alongside the image.
DoP Standard and DoP Preview accept text alone or text + image. Kling v2.1 Pro and Seedance v1 Pro are image-to-video only — without a valid hosted image URL they have nothing to animate.
model must be one of the four known values; an unknown value raises before any network call.
duration is sent in seconds as an integer (1–15). Individual guest models may clamp to their own supported lengths server-side.
aspect_ratio passes straight through to the API.
Pick the right node
Nebula exposes one Higgsfield node. The choice you make is the model param, not the node.
Nebula node (id / display)
Model selected (model value) → endpoint
Capability
Key params
higgsfield / "Higgsfield"
higgsfield-ai/dop/standard → POST /higgsfield-ai/dop/standard (default)
Text-to-video and image-to-video, flagship DoP fidelity
DoP Standard — default; best quality; use for final renders and any pure text-to-video shot.
DoP Preview — iterate quickly on prompt/camera/mood, then switch to DoP Standard for the final.
Kling v2.1 Pro / Seedance v1 Pro — reach for these only when you have a real hosted image and want that model's specific motion character. Don't select them for a text-only graph; with no image URL they have nothing to work from.
Param reference
Single node higgsfield (category video-gen). Ports and params are grounded in backend/data/node_definitions.json and the handler.
Input ports
prompt — Text, required. The scene/motion description. Sent as prompt.
image — Image, optional. Sent as image_urlonly when it's an http(s) URL (see rule 5). For Kling/Seedance this is effectively required.
Output ports
video — Video. Local path to the downloaded .mp4 after completed.
Params
Param key
Type
Default
Range / enum
Notes
model
enum
higgsfield-ai/dop/standard
higgsfield-ai/dop/standard ("DoP Standard (T2V + I2V)") · higgsfield-ai/dop/preview ("DoP Preview") · kling-video/v2.1/pro/image-to-video ("Kling v2.1 Pro (I2V)") · bytedance/seedance/v1/pro/image-to-video ("Seedance v1 Pro (I2V)")
Selects the engine and the submit path. Unknown value → handler raises.
duration
integer
5
1–15 (seconds)
Sent as duration int. Guest models may clamp server-side.
aspect_ratio
enum
16:9
16:9 · 9:16 · 1:1
Passed straight through. 9:16 for vertical/Reels, 1:1 for square.
Recipes
All recipes use the real node id higgsfield.
Pure text-to-video (one node). Drop a higgsfield node, leave model on DoP Standard, set prompt to e.g. "slow dolly-in on a neon-lit Tokyo alley at night, rain, cinematic", duration 5, aspect_ratio 16:9, leave image empty. Run → one MP4 on the video port. No image input needed.
Image → motion (animate a generated still). Use an image node that emits a hosted URL — e.g. a gemini or gpt-image-2 (FAL-backed) image node whose output is a public image URL — and wire that into the higgsfield node's image port. Switch model to Kling v2.1 Pro (I2V) or Seedance v1 Pro (I2V), set prompt to the motion ("camera orbits the subject, hair blowing in the wind"), aspect_ratio 9:16. Confirm the upstream output is a URL, not a local path, or the image is silently dropped and you get a text-only render.
Templated prompt from a Text node, fast then final. Feed prompt from an upstream Text/LLM node so you can template or batch the scene description. Keep image empty. Set model to DoP Preview to iterate cheaply on wording and camera, then flip model to DoP Standard for the final high-fidelity render. Same node, same wiring — only the model param changes.
In the nebula_nodes context
Node id:higgsfield. Display name: "Higgsfield". Category:video-gen (same palette group as the other text/image-to-video generators).
Ports: in prompt (Text, required) + image (Image, optional URL); out video (Video).
Chaining rules:
Anything that emits text (Text/LLM node) → prompt.
Anything that emits a hosted image URL → image. A node emitting a local file path will be ignored by this handler (no auto-upload). Prefer an upstream node whose output is a public/served URL.
video output chains into any node that consumes Video (e.g. a downstream video-processing or save node).
How output renders: on completed, the handler downloads the MP4 from video.url into the run dir and returns {"video": {"type": "Video", "value": "<local path>"}}. The canvas serves/plays it like any other video output.
Capability boundaries (what Higgsfield's API can do that this node does NOT expose)
Do not promise these through the higgsfield node — none are wired (from the audit gap table in docs/api-guides/higgsfield.md; roughly ~25% of the Higgsfield API surface is exposed, video/I2V only):
Motion / camera presets (getMotions(), motion-id + strength) — DoP's signature 50+ camera-move recipes are not surfaced. Camera motion is prompt-only in Nebula.
Soul text-to-image (higgsfield-ai/soul/standard, reve/text-to-image) — Higgsfield's flagship image model is not wired; Nebula uses other providers for images.
Soul styles & SoulID character references (getSoulStyles(), createSoulId(), style_id/style_strength, custom_reference_id) — no style presets or character-consistency controls.
Speak — audio-driven talking-head / lipsync video (/v1/speak/higgsfield, image + audio → lipsync). Not exposed.
Request cancellation (POST /requests/{id}/cancel) — the submit response includes a cancel_url. As of 2026-06-05, a poll cancellation fires a best-effort POST to that cancel_url so the in-flight job stops upstream instead of running to completion (skipped if the submit response carried no cancel_url).
Direct image upload (uploadImage / upload) — the node only accepts an existing http(s) image URL; it cannot upload a local file.
Webhook delivery (hf_webhook / webhook_url) — Nebula polls instead (expected for this architecture).
Broader model catalog — the same POST /{model_id} pattern can reach any model at cloud.higgsfield.ai/explore, but Nebula hard-codes only the 4 models above.
If the user needs any of these, the honest answer is: the Higgsfield node is video-only (text/image-to-video). They'd need a different Nebula node for images/audio, or the node/handler would have to be extended.
