SeeDream Generate Image

SeeDream Generate Image is the node that gives your local ComfyUI a window into ByteDance's flagship image model - Seedream - without you needing a GPU that can run it. "SeeDream" is a romanization of the Chinese for Seedream; under the hood this node calls the Doubao Seedream API on Volcengine Ark and returns the result as a normal ComfyUI IMAGE tensor, so it flows into the rest of your graph like any locally-generated image.

It's a cloud node, full stop. Nothing runs locally except the request. That's a feature if you want Seedream 4.0's quality on a potato of a machine, and it's a non-starter if you wanted everything offline. Seedream is one of the closed API models the community constantly measures open models against - ByteDance's two-track posture, open components alongside a fully closed flagship line, is well documented. This node is the bridge to the closed half.

The inputs that matter

• api_key - your Ark API key (the tooltip literally calls it ARK_API_KEY). Grab one from the Volcengine/Ark console. There's no key file setup for this node - it goes straight in the field.
• model - defaults to doubao-seedream-4.0.
• prompt - the prompt. This model family is a Chinese prompt powerhouse, but English works fine.
• width / height - 1024 to 4096, step 64, default 2048 on both. It's a big-model workflow; these defaults are not typos.
• sequential_image_generation - a mode for generating a sequence of images (think storyboards); pair it with max_images (1–10) to cap how many.
• response_format - url or base64. If your network can't reach the returned download URL (or you want to avoid a second fetch), base64 returns the image inline.
• retry_count, timeout - retries on failure and request timeout (0 = no timeout).
• region - cn-beijing by default; leave it unless your Ark console points elsewhere.
• proxy_url - the escape hatch: protocol://user:pass@addr:port, supporting http, https, socks5 and socks5h. You'll know if you need this.
• seed - -1 for random.
• extra - a JSON string for additional request parameters beyond what the node exposes.

Optional inputs: image for plain img2img, and image1 through image4 for multi-image workflows (reference sets, style transfer, that sort of thing). Outputs are image (the IMAGE tensor) and response_text (whatever the API returned alongside the image).

Install and setup

Same pack as the rest of Yogurt Nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/yogurt7771/ComfyUI-YogurtNodes.git
cd ComfyUI-YogurtNodes
pip install -r requirements.txt

or ComfyUI Manager → ComfyUI-YogurtNodes → install → restart. It's under Yogurt Nodes → LLM. No model files, no heavy deps - just requests for the API call and your key.

The gotcha people actually hit

It's regional. The Ark API's availability has been effectively limited to Asia - there are threads from North American users asking exactly "can I even get a ByteDance API key out here?" and the answer tends to be a shrug or a proxy. Before you build a whole pipeline on this node, check that your region can reach the API at all; if it can't, the proxy_url field is the intended workaround. Also expect this to be a billed API - you're paying per generation, not running free local inference. Read your key's quotas before you loop it for a batch of a hundred frames.