GLM Vision BBox Dual Extractor

The single-image GLM Vision BBox Extractor sends one image to GLM and waits. This node runs two of those calls at once - two images, two prompts, even two separate API endpoints and keys - and returns both results. If your pipeline needs to box up two regions of the same product shot, or two different products in one pass, this halves the wall-clock wait.

What's different from the single extractor

Same underlying request and cleaning logic (urllib, OpenAI-compatible chat-completions, JSON-only output with desc/class/bbox preserved verbatim). The structure is just doubled:

• image_1 + prompt_1 → endpoint_1 + api_key_1
• image_2 + prompt_2 → endpoint_2 + api_key_2
• model is shared by both calls (default glm-4.6v-flash).

The two requests run concurrently via a thread pool, and the outputs come back in slot order - bbox_json_1, bbox_json_2 - even if the second request finishes first. Both inputs are validated before either request starts, and if either call fails, the whole node errors with no partial output. That's deliberate: it keeps you from silently building a mask from one good and one failed result.

Why you'd use it

Two of the endpoint/api_key pairs means you can point the second call at a different GLM model, a proxy, or even a different vendor's OpenAI-compatible endpoint - useful if one region's prompt needs a stronger model. Or just run one prompt for banners and another for text on the same pair of images. If you only need one call, the single extractor is simpler; this node pays for itself when two independent boxes have to happen anyway.

Installing

Part of comfyui-dsocr-bbox. ComfyUI Manager (search "dsocr") or:

cd ComfyUI/custom_nodes
git clone https://github.com/maomaozi/comfyui-dsocr-bbox

restart. The GLM nodes need no extra packages - stdlib HTTP only. The pack's rapidocr-onnxruntime requirement is for the local-OCR nodes, not these.

Gotchas

Two things bite. First, each image input must be exactly one image - batches error out, same as the single extractor. Second, remember the concurrent calls both consume API capacity: two simultaneous requests to the free Flash tier can trip rate limits on both at once, so if you're seeing 429s on a busy run, the retry-with-backoff handles it, but it's still worth budgeting for. And as always with these nodes, pin the coordinate format in your prompts - the extractor preserves whatever GLM returns, so coord_base/coordinate_order downstream must match what you asked for.