Seedream Image Generate With Web Search

Every image model has the same failure mode: it only knows what its training data knew. Ask it to draw something specific and current - a landmark's new renovation, a product that launched last month, a niche subject it was never fed - and it hallucinates with total confidence. SeedreamImageGenerateWithWebSearch is the node for exactly that situation. It wraps the same Seedream image API, but before generating it lets the model actually search the web, so the image is grounded in what's real rather than what the checkpoint remembers.

The mechanism is simple and it's all in the request. This node is pinned to doubao-seedream-5-0-260128 - a model version that supports a web_search tool - and when you flip enable_web_search on, it appends a ContentGenerationTool(type="web_search") to the API call. Turn it off and it's a plain Seedream 5.0 image node. That's the whole difference, and it's the entire reason this node exists as a separate class instead of a checkbox on the base one: the tool is only supported by that specific model, so the node hard-pins it for you rather than letting you shoot yourself in the foot with a model that can't search.

What you set

It's the base node's input list, minus the model picker. You get enable_web_search (default on - it's the point), prompt, aspect_ratio with the full preset list, sequential_image_generation + max_images + stream for multi-image, response_format, watermark, use_local_images, seed, and enable_auto_retry. Plus the five optional image inputs, because searching the web doesn't stop you from also handing it reference images. Outputs are the standard pair: images as IMAGE tensors and the text log.

When it's worth it

Web search is a tool, and tools cost tokens. For a generic "cyberpunk city at night" you're paying extra for nothing - the model knows that cold. Use it when the prompt hinges on facts: "the new main hall of X station," "a still-life of the 2026 season's signature product," anything where the model either knows the answer or will confidently invent one. This is also the node to grab for image-to-image with a factual twist, since the search grounds the whole generation, not just the text part.

The trade-offs worth knowing: you're stuck on 5.0 for this variant (if you need 5.0 Pro quality, you're on the base node without search), and it's a cloud call like every node in this pack, so a web search on top of generation means a slower, slightly pricier request. The text output tells you what actually got searched and generated, so read it before assuming the model silently used the web.

Install

Same pack, same steps. ComfyUI Manager → search "Seedream Image Generate", or:

cd ComfyUI/custom_nodes
git clone https://github.com/dzy1128/Seedream-Image-Generate-ComfyUI
pip install volcengine-python-sdk[ark]

Set ARK_API_KEY, activate a Seedream model in the Volcengine console, and make sure the machine running ComfyUI has outbound network access - a node that searches the web is useless behind a firewall. No model files, no local dependencies beyond the SDK. If you've already got the pack installed, this node was there all along; it just lives a few pixels down in the same image/generation menu.