Zhenzhen Image GK v2 图像编辑（1-3 图）

Editing images by instruction rather than by mask is where the image world moved, and Zhenzhen_Image_GK_V2_Edit is that idea inside the T8mars/ComfyUI_Seedance pack: connect one to three reference images, type what you want changed, and the hosted GK v2 edit model returns an edited image. No mask painting, no inpainting setup, no denoise dials - the instruction is the edit.

This is the same philosophical shift the KB's inpainting essay tracks: between 2025 and 2026, instruction-editing models took over most of what people used masks for, because they fix things from a sentence. The trade, and it's a real one, is that the model sees the whole image and can change more than you asked - drift on the parts you wanted left alone. If you need pixel-exact "change only this region and nothing else," that's what the family's dedicated segment/region-edit nodes (or a local masked-inpaint workflow) are for. This node is the fast, conversational path.

How it works

Connect image1 through image3 (the API strictly supports 1 to 3 reference images), write the edit instruction in prompt, and the node handles upload and the async submit → poll → download cycle on the reseller's /v1/image/generations endpoint. Two controls shape the output canvas: aspect_ratio (14 choices including auto, with wide formats like 19.5:9 and 20:9 in the list) and resolution (1k or 2k). n requests 1 to 10 results from the API - and as with the other GK v2 nodes, only the primary result gets downloaded, so treat n as a credit toggle.

There's also an nsfw_check boolean that enables the documented content check. On by default it is not - the default is off, so you're opting into stricter filtering if you flip it on. That's the reverse of most moderation controls and worth knowing.

Inputs and outputs that matter

• prompt - the edit instruction. This is the whole job; be specific.
• image1–image3 - reference images, 1 to 3.
• aspect_ratio - auto plus 13 fixed ratios, including ultra-wide formats.
• resolution - 1k or 2k.
• n - 1 to 10, downloads the primary result.
• nsfw_check - content-check toggle, default off.

Outputs: image, image_url, task_id, response.

Install

One node in T8mars/ComfyUI_Seedance:

# ComfyUI Manager: search "ComfyUI Seedance", or:
comfy node install seedance

Or clone https://github.com/T8mars/ComfyUI_Seedance.git into ComfyUI/custom_nodes, pip install -r requirements.txt, restart, and configure a key from api.seedance.nz/console via the Seedance API Config node or SEEDANCE_API_KEY.

Where people get burned

The classic instruction-edit failure is the model editing everything: "remove the coffee cup" can also subtly re-light the whole scene. Look at the full response if the result surprises you, and re-run with tighter language rather than fighting it. Also don't confuse this with the region-edit node - that one (with its own page) edits only a selected region of a segmented image and needs an image_id from the segmentation step. If you want only the region changed, that's the node; if you're fine with a whole-image edit, this is the one. And remember the hosted-model rule: your reference images leave your machine, so don't feed this node anything you need to keep private.