并发提交｜Zhenzhen Image GK v2 区域编辑

Most edit nodes make you describe the change and hope the model localizes it. Zhenzhen Image GK v2 区域编辑 is the surgical alternative: you tell it where to edit - by object index, by bounding box, or by an explicit region - and only that area changes. In its concurrent submit form, it lets you run a batch of these targeted edits in parallel. It's also the one node in the pack that can't stand alone, which is exactly what makes it interesting.

The inputs are a chain, not a settings panel:

• image_id - required, and it has to come from the GK v2 segment node (Zhenzhen Image GK v2 智能分割), which first decomposes your source image and returns this ID. You can't feed this node a raw image; it edits a segmented image that the API already knows about.
• selection_mode - how you describe the target region: object_indices, boxes, or selection_regions.
• selection_json - a JSON list matching the mode. Object indices are a list of non-negative integers ([0] is the default); boxes are lists of numeric coordinates; selection regions are a list of objects. The node validates the JSON shape against the mode before anything is submitted, so malformed input fails fast with a readable error rather than a cryptic API response.
• prompt - the region editing instruction.

Output is the single future (SEEDANCE_IMAGE_FUTURE) into 并发接收图片（30 路）.

How it works

The concurrent submit machinery is unchanged - preflight, thread-pool handoff (up to 30 workers, SEEDANCE_IMAGE_CONCURRENCY), future returned - but the underlying request is special: it references the segmentation result by image_id and sends the selection structure the API documents. That's why the pipeline needs two hops (segment → region edit), and the README's release notes say this node family supports all three selection structures. In a batch context, one segment pass can feed several region-edit slots, each targeting a different object in the same image - which is genuinely useful for "change the shirt, leave the face, swap the background" style workflows done in one run.

Installing it

ComfyUI Manager → search ComfyUI Seedance, or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/ComfyUI_Seedance.git
python -m pip install -r ComfyUI_Seedance/requirements.txt

Restart, look under Seedance/并发提交. requests and an api.seedance.nz key; nothing to download.

Where people get burned

• No image_id, no edit. This is the classic trap: someone connects an image expecting magic and gets a "must connect image_id" error. Segment first, then edit. The segment node isn't one of the twenty you'd google separately, but it's the mandatory upstream for this one.
• Wrong JSON shape for the mode. object_indices wants integers, boxes wants numeric coordinate lists, selection_regions wants objects. The node's validator catches mismatches, but you'll save time by matching the docs from the start.
• Thinking it's a general-purpose edit node. It's a region editor. For whole-image restyles you want the GK v2 edit node; this one exists for targeted changes.

Standing fine print applies - metered per render, expiring links, data leaving the machine. If you're doing object-level editing at batch scale, though, this node plus the segment node is the pack's most capable two-step, and the concurrent version is why it scales.