图片｜Grok Imagine Image 参考图/修图

This is the reference-image/editing sibling of the pack's plain Grok Imagine text-to-image node - same grok-imagine-image model, same ~0.208 RMB per image, but instead of "make something from nothing" it takes one to three reference images and redraws them to your instructions. Think of it as a cheap, hands-free retouch: swap a background while keeping the subject, drag a product into a new scene, push a mediocre render up to "commercial product shot" quality. For anyone doing e-commerce content - which is exactly who this pack is aimed at - that's the daily grind.

It's still a closed-model API call through the Tikpan reseller, so the same ground rules apply as every node in this pack: your images leave your machine, you pay per call, and the model's moderation is what it is. Editing a photo of a person with a closed service is where you'll feel that filter fastest.

How it works

Under the hood it POSTs your prompt plus the reference image(s) to Tikpan's /v1/images/edits endpoint, waits, and returns the edited result as an IMAGE tensor. 参考图1 is required and is the "what to change" anchor; 参考图2 and 参考图3 are optional style/fusion references (the node caps at three). The single most important thing you control is 编辑指令 - the prompt that tells the model what to keep and what to move. The default is decent ("keep the main subject consistent, improve composition, lighting, details"), but this is where you actually earn the output: be explicit about what must not change.

A practical pattern that comes straight from the pack's own docs: when you have multiple reference images, name each one's job in the prompt - "image 1 keeps the person, image 2 is the outfit reference, image 3 is the background." And when something is critical (a logo, a face, a product's proportions), say it must be preserved.

The inputs that matter

• 参考图1 - required, the base image being redrawn.
• 编辑指令 - the instruction text; English works better than Chinese with these models.
• 画面比例 - auto follows the reference image, which is usually what you want for edits.
• 清晰度 - 1k for quick drafts, 2k when you're actually shipping the file.
• 生成张数 - more results = more billed images. Start at 1.

Optional: 跳过错误 gives you a black image instead of a hard crash when the API fails (good for batch retouching pipelines), and 返回方式 can stay on the default cloud URL.

Outputs are 编辑结果 (IMAGE) and 渲染日志 (STRING). Save/preview the image like any other node output; the log carries the raw API response when something goes wrong.

Install and first run

Same as the rest of the pack - nothing special to install, no model files:

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

Restart ComfyUI, paste your sk- key from https://tikpan.com into the API_密钥 field, and connect a Load Image to 参考图1. That's it - the code is plain requests + Pillow on ComfyUI's stock deps.

Common issues

The failure modes are mostly upstream. A 401 is a bad key, 402 is an empty balance, 429 means you're calling too fast. Beyond that, the two you'll actually hit: upload size - if a reference image is huge, compress it or drop the resolution first, and prefer a hosted URL over a giant local file; and the model ignoring part of your instruction - that's usually a prompt that didn't name what to preserve. Also worth knowing: if the request errors, check the log before retrying, because a failed call can still leave a charge on the account.