Doubao Seedream ImageGen API Node

Every image node in this pack returns a URL string that you then have to convert. Not this one. Doubao Seedream ImageGen API Node is the exception that makes you feel spoiled: it talks to ByteDance's Seedream image API on Volcano Ark and hands you a ready-to-preview IMAGE tensor plus the raw response string. If you're building a workflow around ByteDance's closed image models - Seedream is the API-only line the community keeps measuring open models against - this is the cleanest entry point in the pack.

Seedream's calling card is strong text rendering, particularly Chinese text, plus an image-edit story mode. If your job involves Chinese signage or lettering that other models mangle, this is a genuinely good reason to have an Ark key.

How it works

The node POSTs a payload with model, prompt, size, and response_format: "b64_json" (so the image comes back inline as base64 - no URL round-trip). It decodes the base64 straight into an IMAGE tensor, batches multiple images if the API returns several, and returns that plus a nicely formatted response_str dump of the full request/response for debugging. It requests watermark: false by default - a small mercy.

Two provider behaviors worth knowing, straight from the source:

• Input image handling. For image-to-image it base64-encodes your optional image input, downscaling large inputs to roughly 1536x1024 first so you don't blast the API with 4K PNGs. A batch input becomes multiple image parts.
• Seedream 4.x quirks. For doubao-seedream-4 models, story_mode maps to a sequential_image_generation field: disabled explicitly disables it, auto turns it on for consistent elements across generations (that's the "Story Mode" the tooltip describes, and it's 4.0+ only). And for doubao-seedream-4-5 models with size "1K", the node silently bumps you to "2K".

The inputs that matter

• prompt - your text. This is a forceInput field, so wire it from a text node rather than typing (typing still works in the widget, it just expects a connection).
• model - default doubao-seedream-4-0-250828; versioned IDs drift, check the Ark console.
• size - 1K / 2K / 4K dropdown.
• story_mode - disabled or auto (Seedream 4.0+).
• image (optional IMAGE) - image-to-image input.
• api_url / api_key - Ark endpoint + an Ark API key from ByteDance's Volcano Engine console.

Outputs: image (IMAGE) and response_str (STRING, the debug/response dump).

Installing it

Part of AhBumm/ComfyUI_BillBum_Nodes. ComfyUI Manager search "billbum", or:

cd ComfyUI/custom_nodes
git clone https://github.com/AhBumm/ComfyUI_BillBum_APIset_Nodes.git
cd ComfyUI_BillBum_APIset_Nodes && pip install -r requirements.txt

Restart ComfyUI. Light deps, no local models, billed per call on Ark.

Where people get burned

• Not a generic key. Ark uses ByteDance's own console and key format; an OpenAI key won't authenticate.
• Failure mode is "empty image" + a string. If something goes wrong, the IMAGE output can come back as None and the details live in response_str - read that string, it's the diagnostics.
• Versioned model names. -250828 style suffixes get superseded; "model not found" usually means update the ID.
• Costs. Text-rendering quality is the reason to pay for this; if you don't need it, local models are free.
• Unlike the sibling video node, this one returns tensors directly - don't add a URL loader after it, you'll be converting pixels to a URL and back for nothing.