PRYX ComfyUI Higgsfield Reference to Video

Character consistency is the oldest unsolved problem in this hobby, and the 2026 answer isn't an IP-Adapter anymore - it's "give the model the reference and let it look while it works." That's what this node is for. It takes ordered multimodal references - images, video clips, audio, even a document or web link, depending on the endpoint - and generates a video that actually uses them: the same face, a specific outfit, a specific camera move.

Fifteen reference-to-video endpoints are in the catalog here: Happy Horse, Wan 3.0 and Wan 2.6/2.7, Seedance 2.0/2.5, MiniMax H3, the Kling Omni and O3 image/video reference variants, Grok Imagine Video 1.5, and Higgsfield's own Cinema Studio 4.0.

How it works

Two ways to feed it. The simple path is the direct sockets: image, video, audio, and end_image where the model supports one - the role is unambiguous, and on this node an image batch gets split into separate ordered references, so a four-image batch is four references, not one.

The other path is references, fed from a Reference Collector. Order is deterministic: direct inputs map first, collector slots follow in numeric order, images/videos/audio/ documents get numbered separately, and batches keep their internal order. The generator validates types, counts and required fields against the selected model before anything uploads - an invalid request never reaches the paid endpoint.

The prompt-token part, which is where this gets interesting

Most endpoints get numbered references delivered in a fixed order and you describe them in prose. Exactly one endpoint in the bundled catalog documents numbered prompt tokens, and the pack says so rather than inventing syntax:

Wan 3.0 · Reference to Video - Image 1, Image 2, Video 1, Audio 1.

There, and only there, you can also use label aliases. Name your Collector slots image_1=person, image_2=outfit, video_1=camera, then write:

Use {{ref:person}} as the main character, {{ref:outfit}} as the clothing,
and {{ref:camera}} as the camera reference.

The pack resolves those to the documented tokens before upload. Unknown labels stop the request; duplicate labels stop the request; and if you write Image 3 when only two images are attached, that stops the request too. On every other endpoint, named aliases are rejected outright - because a silently-ignored alias is how you pay for a render of the wrong person. Reference Preview shows you the mapping and the resolved prompt first, for free.

The inputs that matter

model is required. Beyond that, prompt (a connected input - wire a text or Config UI Prompt node in), duration (union shows 5, 10; individual models run to 30), resolution (720p/1080p/480p/4k/2K), aspect_ratio (ten options including adaptive), generate_audio, enable_thinking, bitrate_mode, seed, mode on the Kling variants, multi_shots with multi_prompt, and aigc_watermark.

Then the cinema knobs, which come from Cinema Studio 4.0 and are the most fun fields in the whole pack: era (1960s–2020s), genre (epic, drama, noir, comedy, horror, action), light (silhouette, practicals, window, overhead-fall, contre-jour, soft-cross), pacing (chaotic, dynamic, calm, single-shot), camera_lens (clean-sharp through halation-vintage), camera_model (modern, 35mm-film, 8mm-film, dv-camcorder), camera_aperture (f14-wide-open, f4-moderate, f11-deep-focus), color_palette (a 50-entry list with names like neon-rain-at-midnight), plus camera_movement and shot_type.

Remember the node shows the union of every model's parameters, and drops fields your selected model doesn't have. Noir lighting settings do nothing on Wan.

Outputs

video (native VIDEO, plus local_file as a path fallback), remote_url as JSON text for logging, request_id (empty on estimates), credits and usd as the provider's estimate with -1 when unavailable, and status - the JSON blob that carries the final reference mapping and provider status. Read status when a render looks wrong; it tells you what actually got sent.

Install

Manager: search PRYX ComfyUI Higgsfield, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/PRYX-STUDIO/pryx-comfyui-higgsfield.git pryx-comfyui-higgsfield
python -m pip install -r pryx-comfyui-higgsfield/requirements.txt

Windows portable uses ..\python_embeded\python.exe. ComfyUI 0.35.0+, frontend 1.53.6+, Python 3.10–3.13, and three small dependencies (higgsfield-client, httpx, jsonschema). No weights to download - the "models" here are a catalog of remote endpoints.

Credentials: Settings → PRYX ComfyUI Higgsfield → Credentials, one key-id:key-secret string, or HF_KEY / HF_API_KEY+HF_API_SECRET for headless installs.

Two habits worth forming, since references make these runs the expensive ones: run Reference Preview after every change to media or labels, and keep max_usd positive so a mis-set duration can't quietly bill you for the 30-second version of your five-second idea.