Graphics Detection Reference

Most nodes in a ComfyUI graph exist to move data toward a generated image. Graphics_Detection_Reference doesn't - it's a standalone diagnostic that benchmarks the GPU (or CPU) it's running on: hardware info, mixed-precision performance, memory bandwidth, operator throughput, and a set of AI-flavored benchmarks (GEMM, tensor core, that kind of thing), with an RTX 4090 comparison baked in as a reference point.

It's a bit of an outlier even inside ComfyUI-WJNodes itself, a personal utility pack (807502278: "a simple node package that I use myself") that's mostly about image cropping, video slicing, and file operations. This one's just a benchmark someone bolted onto their own node pack because it was convenient to have.

How it works

Run it once, and it executes whichever category of tests you picked and prints the results as text - this is an output node, meaning it's meant to be a terminal step you actually look at, not something you wire further downstream. test_items controls the scope: basic for a quick hardware read, precision/memory/operator/ai for narrower deep-dives, or all to run the full suite. test_batch sets how large a batch the tests use internally - the description flags that small batches (like 3) run fast but produce less stable numbers, so treat a low-batch result as a rough estimate rather than a precise one.

The inputs and outputs that matter

• test_items (default all) - which category of test to run. Start with basic if you just want to confirm your GPU is detected correctly before running a heavier suite.
• reference_gpu (default "Other/To be added") - pick a GPU to compare against, with several named options including an RTX 4090 baseline; some entries in the list are explicitly marked as not-yet-filled-in reference data.
• test_batch (default 20, range 1–4096) - the batch size the tests use. Small batches are quick but noisier.
• device (default cpu) - which device to run the benchmark on.
• language (default English) - switches the output text between English and Chinese.

Output is a single info string - the benchmark report, meant to be read directly rather than parsed by a downstream node.

How to install it

Via ComfyUI Manager, search "ComfyUI-WJNodes." Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/807502278/ComfyUI-WJNodes.git

Restart ComfyUI. No model downloads needed - the benchmark tests run directly against your compute stack, not a downloaded checkpoint.

Common issues & troubleshooting

Results feel inconsistent between runs. That's expected at low test_batch values - the node itself warns that small batches produce faster but less stable numbers. Bump the batch size up if you want a more repeatable read, accepting the longer run time that comes with it.

Not sure what to compare your numbers against. The reference_gpu list includes several placeholder entries explicitly marked "to be added" - treat those as aspirational rather than filled-in comparison data, and lean on the RTX 4090 entry, which does have real numbers behind it.

Run takes a long time on all. That's running every test category back to back - switch to a single category like basic or memory if you just need one number quickly rather than a full report.

No community benchmarks to sanity-check against. This node has essentially no public discussion, so there's no crowd-sourced table of "what a 4090 should score" to compare your run to beyond what's baked into the node itself. Treat it as a relative, before/after tool on your own hardware rather than an authoritative external benchmark.