TensorRT Auto Converter 🔧

Let me save you the hour I just spent: this node does not convert anything to TensorRT. I don't say that casually - I checked the source, and the code literally contains the comment "This is a placeholder for actual TensorRT conversion... For now, we'll simulate the conversion process." The trt_model output is the exact same model you put in, passed straight through. You will not get a 40–80% speedup. You will get a nicely formatted status report describing the speedup you're not getting.

This is the most important thing to understand about the whole comfyui-rtx4090-nodes pack: its flashiest node is a simulation. Wire it into a sampler and your workflow runs exactly as if it weren't there - which means it's harmless, but also that it does nothing.

What it thinks it does

Feed it a model, pick precision (fp16 / fp32 / int8), decide static_shape, batch_size and optimize_for (speed / memory / balanced), optionally set width/height, and the conversion_report output narrates the whole imaginary process: "Building TensorRT engine... ✅ Engine built successfully!" plus a speedup estimate lifted straight from the README's table (fp16 40–60%, int8 60–80%, fp32 20–30%).

The trt_model MODEL output is the input model, untouched. There's no TensorRT engine, no ONNX export, no TensorRT dependency anywhere in the pack's install. Even the int8 path - which in real TensorRT requires a calibration step over representative data - is never mentioned again after the dropdown.

What to do instead

If you actually want TensorRT acceleration in ComfyUI, the real answer is NVIDIA's own ComfyUI-TensorRT plugin. That's the project the community actually talks about: people report roughly 1.5–2x on supported models, with the usual caveats (engines are shape- and model-specific, LoRA support has been flaky, and it's a build that breaks fairly often). Real TensorRT is powerful, but it's a serious commitment - not something a single auto-converter node in a tiny pack is going to give you.

Is there any reason to use this node?

Sort of, and it's a strange one. Because trt_model passes the model through untouched, you can drop the node into a workflow, leave all defaults alone, and use conversion_report purely as a display of what could happen - like a note-to-self taped to the graph. If that's not your style, the honest move is to skip the pack entirely for TensorRT purposes and either use the NVIDIA plugin or just keep sampling the plain model, because ComfyUI's standard inference is already decently fast on a 4090 without engine conversion.

Install

Same as the rest of the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/joe002/comfyui-rtx4090-nodes.git

Restart ComfyUI, or grab it via Manager under "comfyui-rtx4090-nodes." No models, no heavy dependencies - which, given the node claims to build TensorRT engines, is itself the tell.

The bottom line, bluntly: it's a stub with a pretty report. It won't break your workflow and it won't speed it up. If a node in this pack gets you excited about TensorRT, let it send you to the real plugin instead.