Debug Pipeline Node

Debug Node checks a LIST. This one checks a MODEL - specifically the pipeline object Model Loader produces. The question it answers is simple but easy to get wrong silently: did the model you asked for actually load the way you asked for it? Right device, right quantization, right task. Model Loader has several optional fields (device_mode, use_bitsandbytes, quantization_type, trust_remote_code) that can each fail to take effect without throwing an obvious error - a cuda request quietly landing on CPU is a genuinely common way for this to go wrong. Better to catch it here than after Caption Generator has been chewing through a hundred images slower than it should.

How it works

Same pass-through shape as Debug Node: feed it model_pipeline, get back debug_model (the same pipeline, unchanged, so the rest of the graph keeps working exactly as if this node weren't there) and debug_text (a readable description of what actually loaded - whatever configuration details the underlying pipeline object exposes about itself).

The inputs and outputs that matter

• model_pipeline (MODEL, required) - the only input, and it has to come from Model Loader.

Two outputs: debug_model (MODEL, pass-through) and debug_text (STRING).

Where to actually put it

Directly between Model Loader and Caption Generator. That's the one useful position - anywhere else and you're checking a model you've already used. Insert it here once when you're first setting up a workflow, or any time you change model_name_or_path, device_mode, or the quantization settings on Model Loader, and confirm the change actually took before you queue real work.

Installing it

Through ComfyUI Manager: search ComfyUI-Transformers-Pipeline, install, restart. By hand:

cd ComfyUI/custom_nodes
git clone https://github.com/mediocreatmybest/ComfyUI-Transformers-Pipeline

then restart. No downloads of its own - it's inspecting whatever Model Loader already pulled down.

Common issues & troubleshooting

debug_text doesn't mention the device or quantization I expected. That's the useful case working as intended - it's telling you the load didn't match your Model Loader settings. device_mode: cuda silently falling back to CPU when no compatible GPU is actually available is the most common version of this; use_bitsandbytes not showing up as applied usually means the quantization backend isn't installed correctly rather than the flag being ignored.

Everything downstream is slow and I don't know why. This is the node to check first, before assuming Caption Generator or your image batch size is the problem. A model running on CPU instead of GPU, or running at full precision instead of quantized, both look like generic slowness downstream - this node is where you actually confirm which one it is.

Nothing looks wrong here, but Caption Generator still misbehaves. Then the model loaded fine and the problem is elsewhere in the graph - check image_batches and the caption-generation settings themselves rather than the model configuration.