图片｜Grok Imagine Image Pro 参考图/修图

Pro editing. If you've seen the standard Grok Imagine edit node, this is it with grok-imagine-image-pro swapped in: one to three reference images in, an 编辑指令 telling the model what to change, and a redrawn image out, billed at 0.728 RMB per call instead of the standard 0.208. The pitch is that Pro holds detail better and blends the reference into the edit more naturally - which matters exactly when you're changing a background but must not touch the product, or restyling a brand asset that needs to stay on-brand.

Honestly, for rough "put this on a beach" experiments the cheap edit node is fine. Pro earns its price on commercial retouching: people photos where the face has to survive the edit, product shots where the label text has to stay legible, logo-inclusive redraws. If it's going on a store page, use Pro. If it's a moodboard, don't.

How it works

Same thin-client mechanism as its siblings: your reference images and prompt go to Tikpan's /v1/images/edits endpoint, the cloud model does the work, and you get back an IMAGE tensor plus a log string. 参考图1 is required - it's the anchor image the model redraws from. 参考图2 and 参考图3 are optional fusion/style references, capped at three total. There's no mask input and no local diffusion involved; the "editing" is the model reimagining the image with your instruction applied, which is why the prompt needs to protect what you care about.

The inputs that matter

• 参考图1 - required, the base image.
• 编辑指令 - where you win or lose. Name the element that must stay identical (face, logo, label, proportions), then describe the change. English tends to track better than Chinese with these models.
• 清晰度 - 1k to iterate, 2k for the final file. On Pro the 2k output is where the detail-holding advantage actually shows.
• 画面比例 - auto follows the reference image, the right default for edits.
• 生成张数 - more variations, more billed calls.

Outputs are 编辑结果 (IMAGE) and 渲染日志 (STRING). Wire the image into Save/Preview like any node.

One habit from the pack's docs worth adopting with multi-reference edits: tell the model what each image is for - "image 1 keeps the person, image 2 is the clothing reference, image 3 is the background." The model uses the extra images a lot better when the prompt assigns them jobs.

Install and first run

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

Restart (or use ComfyUI Manager, search "Tikpan"), paste your sk- key from https://tikpan.com, and connect a Load Image to 参考图1. No model downloads, no extra pip packages - the pack runs on ComfyUI's stock requests/Pillow stack.

Common issues

The 401/402/429 key-and-quota errors you see everywhere in this pack apply here too. On top of that, big reference images are the classic upload failure - compress first, and if your image is already hosted somewhere public, that URL route is more reliable than uploading a 20MB local file. And because each failed Pro edit still might draw from your balance, read the 渲染日志 before you blindly re-run a failed call.