Zhenzhen Doubao Seedance 2.0

Let's get the honest version out front: this node does not run a model on your machine. Comfly_Doubao_Seedance2_0 is a paid API client. You feed it a prompt (and optionally reference media), it ships the job to Doubao's Seedance 2.0 video model through the Comfly relay, waits for the render, and hands you back a video. No checkpoint downloads, no VRAM bar, just a per-generation bill. If you want Seedance-quality video without owning a 5090, that trade is the whole point of the Comfly pack.

Where this node earns its keep is references. Text-only works, but the fun starts when you wire in a first_frame image and a prompt like "the character from the first frame walks toward the camera" - Seedance 2.0 is genuinely good at character-consistent video. Beyond that you get last_frame for closing a loop, up to nine ref_image slots for style/subject grounding, three video references, three audio references, and a generate_audio toggle (default on) so Seedance produces sound instead of silence.

The inputs that actually matter

• prompt - multiline, required. Write it as a scene description, not a Midjourney one-liner.
• model - doubao-seedance-2-0-260128 or the -fast variant. Fast is cheaper and worse; standard is the default for a reason.
• duration - 4 to 15 seconds, step 1.
• ratio - 16:9, 9:16, 1:1, 4:3, 3:4, 21:9, 9:21, or adaptive.
• resolution - 720p, 480p, or native1080p. Native is what you want for final renders; 720p keeps the bill down while you iterate.
• return_last_frame - flip this on and you get the final frame out as an image. Handy for chaining scenes or seeding the next video.
• watermark - default off, so you're not stuck with one.

Outputs: video (the VIDEO object - the README explicitly confirms these video outputs plug straight into ComfyUI's native Save Video node), plus task_id, response, video_url, and last_frame_image for the debug/asset side of things.

The asset bundle path

For heavy multi-reference jobs there's a newer, cheaper route: upload your media once with Comfly_Doubao_Seedance2_0_Asset (you get stable asset_ids back), bundle them with Comfly_Doubao_Seedance2_0_AssetIdBundle, and feed that into this node's asset_bundle input. It avoids re-uploading (and re-paying for) the same files on every run - the pack even ships a "seedance2.0 上传素材ID 多参工作流.json" example workflow for exactly this.

Install and key

Same as every Comfly node - grab the pack once and it's shared:

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

Restart ComfyUI, or just use ComfyUI Manager and search "Comfyui_Comfly". There are no models to download because nothing runs locally. You need an account and credits at https://ai.comfly.chat - put the key in Comflyapi.json (or the Comfly_api_set node) and every node in the pack reads it. The per-node apikey input is just an override.

Where people get burned

It's a paid, async call that takes minutes - a run isn't free retry practice. The good news: failed tasks refund, and the relay keeps async task logs at ai.comfly.chat so you can see exactly what the API said instead of guessing. And since this is a third-party relay, outages happen; that's what the hk/us fallback lines in Comfly_api_set exist for. Author only tests on Windows 11, so if you're on Linux or macOS and something odd happens, issues page is the way in.