🫎NIUNIU API-即梦4.5

Quick honesty up front: this is the same node as the pack's "大香蕉" one, with the model name swapped out. The source is literally a subclass - NewApiJimeng45Node inherits everything from NewApiBanana2Node, overrides the default model to jimeng-4.5, and calls it a day. So if you've read that article, you know 90% of this one. What's left is worth saying: this is how you get 即梦 (Jimeng), ByteDance's closed image model, inside a ComfyUI graph without any local weights.

What it's for

Jimeng 4.5 is ByteDance's current image-generation model, the same family behind their Dreamina/Jimeng consumer app. It has no open weights - you cannot download it, so the only door into ComfyUI is an API. This node is that door. You type a prompt, it sends it to a New API style relay, and the returned image drops onto your canvas as a normal IMAGE tensor. Same tradeoffs as every wrapper node: per-call cost, your prompt (and reference images) leaving the machine, and moderation enforced at ByteDance, not by anything you can configure here.

One thing this node is not: it doesn't do anything special with Jimeng's quirks. It sends a model name over HTTP like any other. If your relay serves jimeng-4.5, great; if it serves a different Jimeng version or a different alias, the 模型版本 field is a free-text box - paste whatever your provider's model list shows.

How it works

Identical mechanics to its sibling: POST to /v1/images/generations (文生图 / text-to-image) or /v1/images/edits (图像编辑 / image editing) with async=true, poll /v1/images/tasks/{id}, then either download the URL or decode base64 into a tensor. The base URL self-corrects if you paste a full /v1/images/generations endpoint. The only code-level difference: the node clamps the batch to your 出图数量 and forces the model name to jimeng-4.5 no matter what you type.

Inputs and outputs

Same field list as the Banana node. The ones you'll actually touch:

• API密钥 - the key from your relay provider. API地址 defaults to https://api.llyapps.com, and any OpenAI-compatible endpoint works.
• 提示词 - the prompt. Defaults to a woman in a forest, real-photography style; you'll replace it.
• 生成模式 - 文生图 or 图像编辑. Edit mode needs at least one 参考图 (reference image) connected.
• 宽高比 (11 presets), 图片分辨率 (1K/2K/4K), 随机种子, 出图数量 (1–4). The 4K option only means anything if the model your relay routes to actually does 4K - the node passes the size through.

Outputs: 图像 (IMAGE tensor, wire to preview/save), 图片链接 (URL string), 任务ID, and 响应信息 (JSON - your debugging friend).

Installing it

The pack ships all five nodes together, so this one comes with the others. ComfyUI Manager: search "ComfyUI-NIUNIUAPI", install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/xiongmaowf/ComfyUI-NIUNIUAPI
# restart ComfyUI

Dependencies are just requests, pillow, numpy - no model downloads, no VRAM, nothing heavy.

Gotchas

The colored 512×512 placeholder convention from its sibling applies here too: black = missing key/address, red = API rejection (401 = wrong key or endpoint; a content block shows up in 响应信息), yellow = poll timeout, white = empty response. The biggest real-world trap is expecting Jimeng's style to be free - it isn't; the relay bills per image, and a 4-image batch at 4K is the kind of session that empties a small balance.

Also worth keeping in your head: this pack is a third-party wrapper that carries your API key and calls the network by design - the category of node that got weaponized once in this ecosystem. It's a small, unremarkable pack, but if you're the careful type, glance at the source before you paste in a key.