Zhenzhen Jimeng Video API

Jimeng - 即梦, ByteDance's video generator, the consumer face of the Seedance line - is one of those closed models you can't run locally and can't easily poke at from ComfyUI. This node fixes the second half. It's a text-to-video (and image-to-video) node for Jimeng through the Comfly proxy, and unlike a lot of API wrappers it hands you the actual generation knobs: duration, aspect ratio, and a cfg_scale you'd otherwise never see.

ByteDance's whole image/video flagship line is API-only - Seedream for images, Seedance for video - so if you want Jimeng-class output in your normal ComfyUI flow, a proxy node like this is genuinely the least-pain route. The README quotes it at roughly 0.6 yuan per 5-second clip and 1.2 yuan per 10-second clip. Per-call, prepaid credits, no GPU involvement.

How it works

Prompt (plus an optional source IMAGE for image-to-video) is sent to the proxy's Jimeng endpoint, the node polls until the video renders, and you get a video output - the URL wrapped in the pack's ComflyVideoAdapter so it wires straight into ComfyUI's Save Video node - alongside the video_url, a task_id for tracking, and the raw response.

The inputs that matter

• prompt - the scene. Describe motion and camera explicitly; Jimeng responds well to clear verbs.
• duration - 5 or 10 seconds. Billing is per duration tier, so a 10s test run costs double the 5s one before you know if the prompt's any good.
• aspect_ratio - six options: 1:1, 21:9, 16:9 (default), 9:16, 4:3, 3:4. Pick the frame for where the clip lands.
• cfg_scale - 0.1–1.0, default 0.5. Lower = more faithful to the prompt but flatter; higher = more model imagination. This is the actual creative dial, and it's the kind of thing most API wrappers hide.
• image - optional IMAGE input. Feed one for image-to-video (animate that still).
• seed - set it to lock a take; 0 means random.

Outputs: video (VIDEO), task_id (STRING), response (STRING), video_url (STRING).

Install

Manager → search Comfyui_Comfly, or:

cd ComfyUI/custom_nodes
git clone https://github.com/ainewsto/Comfyui_Comfly

Restart. No model downloads - the requirements.txt is the usual aiohttp/requests/Pillow/numpy set plus the torch/transformers already in your environment. Key from ai.comfly.chat in Comflyapi.json or the node's api_key field.

Where people get burned

Duration is the sneaky cost lever - 10 seconds is exactly double the price and probably triple the wait, so don't default to it. And like every video node in this pack, Jimeng generation is a cloud job with a real queue: if the run fails partway you'll see it in response (check the task_id in the proxy's async-task list), not as a nice node error. Keep your prompts cloud-safe - you're sending them through a third-party proxy to a ByteDance service, and the hosted video links you get back aren't permanent storage.