Zhenzhen Image GK v2 智能分割

Zhenzhen_Image_GK_V2_Segment is the odd one out in T8mars/ComfyUI_Seedance: it doesn't generate anything. It takes a completed GK v2 image-generation task and asks the server to segment it - split the picture into its objects, give each one an ID, and hand back structured JSON. It's a processing node, the middle step of the pack's segment → region-edit pipeline, and it's the kind of thing you'd otherwise need a local SAM/BiRefNet or an ADetailer-style detector stack to do. Here it happens in the cloud on the same task that made the image.

The mechanism is worth spelling out because it changes how you wire the workflow. Because segmentation runs on the server side against the generation task, this node doesn't take an image - it takes a source_task_id, which you get from the task_id output of a completed single-image GK v2 generation node (e.g. Zhenzhen_Image_GK_V2). You can't feed it an arbitrary local image; it segments what the API already generated. That's the "walled garden" nature of this family: image, segment, and region-edit are all hosted operations on one task chain.

How it works

Connect source_task_id and set include_mask_rle if you want mask RLE data in the result (default off). The node submits a segment operation, polls it to completion, and returns:

• image_id - the ID that feeds Zhenzhen_Image_GK_V2_Region_Edit.
• objects_json - the detected objects as pretty-printed JSON.
• result_json - the full segmentation result.
• task_id / response - the operation's own task ID and raw response.

This node is not an output node - it has no OUTPUT_NODE = True in the class - which in practice means you're expected to pass its outputs forward to the region-edit node rather than treat it as a terminal. The whole design is "segment now, edit later."

Inputs and outputs that matter

• source_task_id - required, from a completed single-image generation task.
• include_mask_rle - include mask RLE data in the result, default off.

Outputs: image_id, objects_json, result_json, 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 source_task_id requirement is where everyone stumbles: there's no IMAGE input, so you can't just load a picture and segment it. If you try, the node's strict validation rejects it - it must be a task ID from this pack's GK v2 generation. Also remember skip_error here returns empty structured outputs on failure rather than a placeholder image, because there is no image. And keep the privacy frame in mind: your generated image is already on their server (that's where it was made), so segmentation adds no new exposure - but it also means the whole segment/edit chain is stuck inside the reseller's ecosystem. If you want to segment arbitrary local images, the local SAM/BiRefNet stack remains the open alternative.