🥟 智绘_豆包 (高清增强版)

Doubao is ByteDance's image-model family, and Seedream is the checkpoint line under it - the one ComfyUI's own Partner Nodes sell alongside Nano Banana. ZH_DoubaoImageGen (🥟 智绘_豆包 高清增强版) is the 智绘灵箱 pack's wrapper for it, defaulting to the doubao-seedream-4-5-251128 model via the VectorEngine endpoint. Same wrapper pattern as the pack's two banana nodes: key in, prompt in, HTTP call, image tensor out. No GPU, no model download - a ByteDance model you can't run locally, delivered into your graph.

How it works

The shape is now familiar: a reused requests.Session POSTs your prompt, seed, aspect ratio, resolution and model choice to base_url (default https://api.vectorengine.ai/v1, so this is the same reseller the banana node uses, just the Doubao product line). Optional ref_image1–ref_image3 sockets send reference images. The "高清增强版 (V21)" is mostly about the resolution handling - it picks the API call's output resolution based on your resolution setting, and it does it intelligently rather than always hammering max res.

Two model choices are on offer: doubao-seedream-4-5-251128 (the default, newest Seedream 4.5) and doubao-seedream-4-0. The resolution dropdown maps 高清 (HD, recommended) vs 标清 (SD, non-fusion mode) - the HD path gets you the good stuff, the SD path is there for the fast/cheap cases. And generation_mode is the interesting one: 多图融合/风格迁移 (multi-image fusion / style transfer) - off by default - versus 多图参考/故事模式 (multi-image reference / story mode) which is auto. In fusion mode you're blending multiple reference images into one result; in reference mode the images steer content but stay recognizable. If you're not using ref images, leave it disabled.

The inputs that matter

api_key is the gate (empty key = nothing). prompt (default "Generate a cute 3D character") is the whole creative job. model picks Seedream 4.5 vs 4.0. Then aspect_ratio (nine choices incl. 21:9 and 9:21), seed (0 = random), and image_quality (JPEG compression, default 95). Outputs are the standard trio: 生成图片 (IMAGE), 详细信息 (STRING), 智绘显示 (STRING) formatted panel. Output node, so plan for it to end a branch.

Install

Part of the 智绘灵箱 (ComfyUI-ZhiHui) pack:

cd ComfyUI/custom_nodes
git clone https://github.com/zhuyungen/ComfyUI-ZhiHui.git

Restart ComfyUI (or ComfyUI Manager, "智绘灵箱" / "ComfyUI-ZhiHui"). Deps: torch/numpy/requests.

Where people get burned

• Same reseller caveats as the banana nodes. VectorEngine proxies ByteDance's API; endpoint availability, rate limits and key issuance are the provider's business. If the model dropdown shows an option that 404s, that's a backend change on their side, not your config.
• generation_mode is not "more images = better." Fusion and reference modes mean different things to the API and behave differently. Leave it on the default (disabled) unless you actually have ref images and a use for them, or you'll get output that ignores the mode entirely.
• Image quality is compression, not resolution. 95 default is fine; pushing to 100 just bloats the file.
• The model id is date-stamped. doubao-seedream-4-5-251128 encodes a 2025-11-28 cut. If ByteDance rotates model versions, the hardcoded dropdown may lag - and the pack's README won't tell you, the provider's docs will.

If you want Seedream in a ComfyUI graph and don't want Comfy's own credit storefront, this is the direct route. The API-wrapper pattern is the same everywhere - key hygiene, endpoint trust, and a prompt away from a ByteDance image on your canvas.