图片｜GPT-Image-2 官方生图

Sometimes the best node is the boring one. TikpanGptImage2OfficialNode is the pack's plain, no-surprises GPT-Image-2 text-to-image node: prompt in, gpt-image-2 via Tikpan's /v1/images/generations, image out. Where the pack surrounds this model with benefit channels, legacy -C compatibility nodes, and a 15-reference multipart editor, this node is the "official" middle lane - the one the author describes as showing only the core parameters Apifox has recorded, which is another way of saying it's the maintained, front-line version. If you want one reliable GPT-Image-2 generation node and don't need multi-reference fusion, this is the default choice.

How it works

Synchronous /v1/images/generations call: payload (prompt, model, resolution tier, aspect ratio, quality, optional output format), decode the returned image, save it locally, emit 🖼️_生成结果图 (IMAGE) plus 📄_渲染日志 (STRING). It's mid-graph (is_output_node: false), so wire the IMAGE into a Save Image or downstream compositor.

The inputs that actually matter

• API_密钥 - sk- from tikpan.com (the header hints at the 0.6 RMB ≈ $1 virtual-token entry, a reminder this is metered).
• 生成指令 - the prompt. Default is a photorealistic aurora scene with "8k超清，电影感" - GPT-Image-2 renders detailed, text-heavy prompts well, so be specific.
• 模型 - gpt-image-2, the only option in this node.
• 分辨率档位 - 512 / 1K (1024, default) / 2K (2048) / 4K (3840). The tooltip is blunt: higher = slower and more expensive. For most web/social use, 1K–2K is the sweet spot.
• 画面比例 - nine options, 1:1 through 9:21 (default 1:1).
• 画质与推理强度 - auto / low / medium (default) / high, mapping to the quality parameter.

Optional: 返回格式 (png / jpg / webp - png default for quality, jpeg for size) and 跳过错误 (on failure, return a black image instead of killing the workflow - handy for unattended runs, but spot-check for black frames).

Installing it

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

restart (or Manager → "Tikpan"). No models, no heavy dependencies - relay + decode/save.

Where people get burned

• It's still a relay call. The clean UI doesn't change the economics: every generation meters, and your prompt leaves the machine for a third-party server. The KB's API-nodes essay is the place to set expectations on both.
• 跳过错误 hides failures. A black image is a silent error. Leave it off while building, on for batch.
• Quality/resolution is the cost dial. 4K + high on every draft is the fastest way to empty the meter. Draft at 1K/medium, re-render the keepers.
• No reference images here. If you need multi-reference or editing, this node won't do it - that's the Gemini 14-ref node, the V2 editor, or the legacy -C pair. This one is deliberately single-purpose.

That's the whole pitch, and it's a good one: the "official" GPT-Image-2 node is the one you can trust to do exactly what it says, with the fewest knobs to mis-set. Sometimes the boring node is the one that earns the install.