Concurrent Submit | zhenzhen-seedream-v5-pro

The concurrent twin of Zhenzhen's Comfly_seedream_v5_pro node - ByteDance's flagship image model, the one that grew out of the Seedream 4.x line into a genuinely strong all-rounder: text-to-image and multi-reference editing, up to 4096×4096 output, and a reputation for character consistency that the pack's own changelog claims beats Nano Banana on people. That last bit is worth a test run if consistency is your bottleneck.

This is the version of the node that takes a raw apikey field (rather than the api_config used by the domestic low-price twin), so it plays fine with the overseas token. And it's the node with the handy image_urls input: a multiline field where you can paste reference image URLs directly, one per line or comma-separated, as an alternative to dragging in image1…image10 tensors. Great for re-running edits off a cached URL list without re-uploading.

Concurrent mechanics: image pool (30 workers), COMFLY_IMAGE_FUTURE task output, Concurrent Collect Images (ComflyConcurrent_Image_Await) collects up to 30 tasks into image_1…image_30.

The inputs

• prompt - the direction. Multiline.
• model - a freeform STRING defaulting to seedream-v5-pro; type other family names here if the reseller adds them.
• size - 1024×1024, 1536×1024, 1024×1536, 2048×2048, 4096×4096, or custom (which uses custom_size, e.g. 2048x1536).
• response_format - url or b64_json. URL keeps the payload light; b64 if you want the image inline in the response.
• output_format - png / jpeg.
• image1 … image10 and image_urls - the two reference routes.
• seed - execution seed (cache-control compatibility note applies).
• apikey and skip_error - the usual. Note this node's skip_error tooltip: "返回空白图" - failure returns a blank image when enabled.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-zhenzhen

Or ComfyUI Manager → search Comfyui-zhenzhen → restart, token into apikey.

Gotchas

• 4096 is a big, slow, expensive file. The pack's cross-border download drama is real for 4K-class outputs - if downloads fail, enable OSS on your token's edit page. Start at 2048.
• Mixed reference routes confuse people. You can fill image1…image10 AND image_urls; if you're using URLs, leave the tensor slots empty. The node has no idea you meant one or the other.
• The blank-image skip mode is dangerous in batches. With skip_error on, a failed slot silently becomes a blank image in your collected batch - great for keeping a pipeline running, awful if you don't check the status output before shipping results.