Higgsfield - Generate Video

Seedance is ByteDance's video line, and like the whole Seedream/Seedance family it is closed - there is no download link, and there never was. So the local-versus-cloud maths is at its purest here: run Wan or LTX locally forever and own the results, or pay per clip for the model everything else gets compared to. HFVideoGenerate is the second option, wired into your graph: prompt or reference images in, a native VIDEO object out, with whatever audio the model generated still attached.

Worth saying before you spend anything: video is where API sessions get expensive, fast. The community's reaction when Seedance first arrived in ComfyUI was mostly about price and subscription lock, and both complaints were fair.

The four routes, and what each one allows

The model dropdown is really four endpoints with different rules, and the node enforces them locally rather than letting the API reject you:

• Seedance 2.5 - Text to Video - prompt only. Attaching references to this route is an error; that's what the other two are for. 4–30 seconds, 480p or 720p.
• Seedance 2.5 - Image to Video - one to two image references, in order: the first is your start frame, a second becomes the end frame. Aspect ratio is derived from the image, so leave the control at 16:9; that value isn't sent. Same 4–30s, 480p/720p.
• Seedance 2.5 - Reference to Video - the flexible one: images, video and audio references guided by your prompt, at least one required. 4–30s, 480p/720p, mp4 or mov.
• Seedance 2.0 - Text to Video - the older, longer-leash route: 4–15 seconds but up to 1080p and even 4k, mp4 only.

The inputs you'll actually set

prompt is the motion description. Required for Text to Video; optional on the other two, where leaving it blank means the field isn't sent - the right move when the start frame already tells the story. duration is whole seconds (4–30, default 5). generate_audio defaults on, and that audio survives into the VIDEO object you get back - the real reason to reach for Seedance over a muted local model. output_format picks mp4 or mov on 2.5 routes and is locked to mp4 on 2.0.

Then the shared controls: generation_id is a local take label, not a seed - change it for a new paid clip, keep it to resume or reuse. timeout_seconds (default 1800, max 7200) is how long you'll wait locally; video polls take a while, so don't shave it. references takes the HF_REFERENCES chain from Reference Images, or from Reference File for a local PNG, MP4 or WAV - the only way to feed video or audio references, since ComfyUI's IMAGE type can't carry them.

Out: video (a VIDEO object - feed it to Save Video, or Get Video Components if you want frames or the audio track), saved_path, and request_id, which you can paste into Resume Request to pull the same clip again later for free.

Install

ComfyUI Manager → search Higgsfield, or:

cd ComfyUI/custom_nodes
git clone https://github.com/w0ver/Higgsfield-api-comfyui-nodes ComfyUI-Higgsfield

Keep the folder landing directly under custom_nodes. Requirements are just requests, so the dependency side is nothing; the setup is the credential. Windows: run Configure API.bat in the node folder (or python custom_nodes/ComfyUI-Higgsfield/configure_api.py) and enter your key ID and secret from open.higgsfield.ai. Linux/macOS: export HF_API_KEY_ID and HF_API_KEY_SECRET in the environment ComfyUI starts from - the encrypted helper is Windows-only, and a .env file next to the code is not read. You also want a recent ComfyUI, since VIDEO and Save Video are recent core additions.

Where people get burned

• "Image to Video needs one start image and optionally one end image" - you either sent zero references or more than two, or one of them wasn't an image. Build the chain with Reference Images and check the order.
• aspect_ratio refused on an Image to Video route. Working as intended - the ratio comes from your image.
• HTTP 422 - almost always duration or resolution out of range for that route (a 20-second 1080p Seedance 2.5 clip, say, or 4k on a 2.5 route).
• 429 or 400 with a concurrency message - you're submitting faster than the account allows. Let existing requests finish; the pack deliberately won't retry a submission that might already have been accepted.
• Node times out mid-poll - nothing is lost. Queue again with identical inputs and generation_id; it resumes the recorded request and downloads then.
• Stopped the queue and nothing came back - ComfyUI's stop button ends your local wait, not the generation. The job is still at Higgsfield, and Resume Request or the console will find it.