CYAI Seedance 视频生成 (cyai.club)

ByteDance's Seedance line has no weights and isn't getting any - the entire Seedream/Seedance family is an API product while ByteDance keeps handing out the components underneath it for free. So if you want Seedance in a Comfy graph, someone has to make an HTTP call on your behalf. That someone is this node. It doesn't run a model; it submits a job to a Volcengine Ark endpoint (through the cyai.club relay by default), polls it, downloads the mp4, and hands you a ComfyUI VIDEO object like a local sampler had made it.

Be clear-eyed: closed model, reseller, per call, your frames leaving the machine. In exchange you get shot quality the open video models still haven't matched - which is the whole reason the API-node category exists.

How it works

The node builds one content array - a text item for the prompt, then an image_url item per frame or reference, each tagged with a role of first_frame, last_frame or reference_image - and POSTs it to /api/v3/contents/generations/tasks. Back comes a task id. It then GETs /api/v3/contents/generations/tasks/{task_id} on a loop until the status lands in succeeded, completed, success, done (or trips failed, error, cancelled, canceled), pulls the video URL out of content.video_url, downloads the bytes, and wraps them with comfy_api's VideoFromFile. The progress bar counts polling rounds, and the sleep between polls is chopped into two-second slices so pressing cancel actually stops the node instead of leaving it hanging.

That download-immediately part matters more than it looks: the video URL is a signed TOS link that expires in 24 hours, and the node fetches it the moment the task reports done, so in practice you never see the expiry. The task list itself is only queryable for 7 days.

The inputs you actually touch

Four fields get you running: api_key (your sk-... from the relay), prompt, and model - a plain text box, not a dropdown, so doubao-seedance-2-0-260128, 2-0-fast-260128, 2-5-260628 and whatever the relay adds next all work without a pack update.

Then pick your mode. Leave first_frame, last_frame and reference_images empty for text-to-video. Wire first_frame (plus optionally last_frame) for image-to-video. Or wire reference_images with up to four images to blend characters or styles. Only one mode at a time - the node raises a ValueError if you mix frames with references, because the API refuses it anyway.

Two defaults worth knowing: generate_audio is on, which surprises people who expected a silent clip, and duration takes 4–30 seconds even though the 2.0 models cap out at 10 and only 2.5 goes to 30. max_wait defaults to 900 seconds; bump it if your relay queues hard at peak.

The output is a single video port. Wire it to ComfyUI's preview node to watch it on canvas, or into Video Helper Suite's save/transcode nodes to get an actual file. If you're doing AI comic work with real faces, also flip upload_images on - more on that below.

Installing it

ComfyUI Manager is the easy path: Custom Nodes Manager, search CYAI Seedance, install, restart. Manual works too:

cd ComfyUI/custom_nodes
git clone https://github.com/Gin3601/comfyui_transit_node.git

and restart ComfyUI.

The genuinely nice part: there is nothing to install. No requirements.txt, dependencies = [] in pyproject.toml, and the shipped code only imports torch, PIL, numpy, comfy_api and stdlib urllib. No pip install, no torch version roulette. You do need a funded key on the relay: the pack ships a usage-query node, except the author commented it out of __init__.py because the balance can't be read through the API at all - check credit in the relay's dashboard.

Where people get burned

ratio must be adaptive in first/last-frame mode. Not a suggestion - the node checks it before submitting and refuses with "首帧/尾帧模式要求 ratio 必须设为 adaptive". If you're doing image-to-video and nothing happens, this is usually why.

Real faces get blocked, and it's not the node's fault. Ark rejects anything it thinks contains a real person: InputImageSensitiveContentDetected.PrivacyInformation. Both base64 and public URLs get scanned, so re-encoding won't help - only an asset:// reference passes. The upload_images toggle automates that whole detour: turn it on, fill the key, keep your images on reference_images or first_frame, and the node uploads them as assets and rewrites the request itself. You don't wire asset_ids when it's on. It does cost you an extra asset upload per image, and it needs at least one image connected or it errors out.

A failed run costs you twice. There's no task-id resume here. If generation succeeds and the download then dies - flaky network, killed ComfyUI - you can't point the node back at that task. You paid, you try again.

Widget positions shift between versions. The author appends new inputs at the end of INPUT_TYPES deliberately: ComfyUI assigns widget values by position, not name, so inserting a field mid-list would slide every value in your old workflow over one slot. Glance at the node after you update.