图片｜豆包图像生成 Seedream

ByteDance's Seedream image line is API-only - the KB's ByteDance panel is explicit that every flagship generation product stays closed while the lab's open releases are the infrastructure parts. TikpanDoubaoImageNode is the relay wrapper that puts Seedream inside a ComfyUI graph: it hits Tikpan /v1/images/generations with your prompt, decodes the response, saves the image, and returns a plain IMAGE batch you can feed straight into a Save Image node. No ByteDance account, no VRAM, no model download. For product-shot work - the pack's whole audience is TikTok e-commerce sellers - this is one of the more practical image nodes in the set.

How it works

Unlike the async video nodes, this is a synchronous call: build the payload (prompt, size tier, aspect ratio, format, watermark toggle, optional reference images as base64), POST to the relay, decode the returned image(s), save them to your ComfyUI output folder, and emit 🖼️_生成图像Batch (IMAGE) plus 📄_渲染日志 (STRING). It's is_output_node: false, meaning it's mid-graph - wire the batch into a save/preview node.

The inputs that actually matter

• API_密钥 - from tikpan.com (this one is on the main relay, not new.ip233.com). The 获取密钥请访问 hint links there.
• 生成指令 - the prompt. The default is a cyberpunk product shot; it handles Chinese and English.
• 清晰度档位 - 2K or 3K. 3K is sharper, slower, pricier. 2K is the sane default for e-commerce.
• 尺寸模式 - 品质档位 (auto-compute from 2K/3K) or 按比例输出精确尺寸 (strict pixel output for your chosen ratio).
• 画面比例 - eight options, 1:1 through 21:9.
• 水印 - 无水印 or 有水印. Doubao's official watermark can be turned on if you want the provenance; off for clean assets.
• 多图生成 - 关闭 (one image) or 自动 (batch candidates, up to 最多生成张数, default 4, max 15). More images = slower + pricier, and 多图失败处理 decides whether a partial failure hard-errors or degrades to the first image.

Optional: 参考图 (an IMAGE batch - up to 14 images) and 图片URL或Base64 (manual URL/base64 lines that merge with the batch).

Installing it

Standard pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/htrert/ComfyUI-Tikpan-Pro

restart, or use ComfyUI Manager → "Tikpan". No dependencies beyond the usual HTTP + PIL stack; nothing heavy.

Where people get burned

• It's metered per image. 3K + multi-image batches burn relay credits fast. The 0.6 RMB ≈ $1 virtual-token pitch in the header hint means entry is cheap, but a 15-image 3K run is not free.
• Reference cap is 14. The tooltip says batch max 14, and the manual URL list merges with it - don't exceed or you'll get a validation error.
• The 跳过错误 pattern. This node doesn't expose it, but the sibling GPT-Image nodes do; here the failure handling lives in 多图失败处理 instead. If a multi-image run keeps hard-failing, drop to 自动降级为首图 to keep the pipeline alive.
• Data leaves the machine. Prompt + reference images go to a third-party relay. Fine for catalog shots, but keep client-confidential assets out of it.

Honest take: Seedream at 2K for product grids, with the 3K tier reserved for hero images, is where this node earns its keep. The watermark toggle and the 14-image reference headroom are the features the e-commerce crowd actually uses.