Load Upscale Tensorrt Model

This node doesn't upscale anything by itself - it's the loader half of a two-node pair. Its whole job is to hand a compiled TensorRT engine to the Upscaler Tensorrt node downstream, the same way a checkpoint loader hands a UNet to a sampler. If you only add one of the two nodes to your graph, you added the wrong one; you need both.

The pack it comes from wraps classic ESRGAN-family upscale models - the "more pixels, no invented detail" kind, not a generative restorer like SeedVR2 - and runs them through NVIDIA TensorRT instead of plain PyTorch, which is where the 2-4x speedup the author claims comes from. That's a real and useful distinction: this thing adds no content, can't hallucinate a face, and does exactly the same job as running 4x-UltraSharp normally, just faster on an NVIDIA GPU.

What you actually set

Two inputs, both dropdowns:

• model - pick from 13 tested ESRGAN-architecture checkpoints: 4x-AnimeSharp, 4x-UltraSharp, 4x-UltraSharpV2_Lite, 4x_foolhardy_Remacri, RealESRGAN_x4, 4x-ClearRealityV1, and a handful of others, defaulting to 4x-AnimeSharp. If you've used any of these with a normal ESRGAN loader before, that's exactly what you're getting here, just TensorRT-accelerated.
• precision - fp16 (default) or fp32. Stick with fp16 unless you have a specific reason not to: these are pixel-restoration models, not identity-sensitive diffusion models, so the fp16 quality hit is negligible and the speed/VRAM win is real.

Output is a single upscaler_trt_model - wire it straight into the Upscaler Tensorrt node.

The mechanism, and why the first run looks stuck

TensorRT doesn't run a model directly - it compiles it into a hardware-specific "engine" ahead of time, baked for your exact GPU and the settings you chose. Since the pack's March 2025 update, that compile happens automatically the first time you run a given model+precision combo through this node; you don't need a separate export step anymore. That first build genuinely takes a while - real reports put it at 20-30 minutes on slower hardware - with no progress bar, which is exactly why people post "is it stuck or downloading?" threads. It isn't stuck. After that first build the engine is cached on disk and loads in seconds.

One consequence worth knowing up front: engines are tied to your GPU and TensorRT version. Upgrade TensorRT later and you'll need to rebuild - that's called out directly in the pack's own known-issues list, not a bug you introduced.

Installing it

Either search "ComfyUI Upscaler TensorRT" in ComfyUI Manager, or do it by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/yuvraj108c/ComfyUI-Upscaler-Tensorrt.git
cd ComfyUI-Upscaler-Tensorrt
pip install -r requirements.txt

Restart ComfyUI afterward. One thing worth flagging before you build a paid pipeline around it: the pack is licensed CC BY-NC-SA 4.0, i.e. non-commercial.

Where people actually get stuck

The most common real failure isn't the node itself, it's pip install -r requirements.txt not fully landing the tensorrt package - especially on Windows portable/embedded builds, where a user hit ModuleNotFoundError: No module named 'tensorrt' on load even after installing via Manager. The community fix that worked for people: run pip against the embedded Python directly with NVIDIA's index -

python_embeded\python.exe -m pip install tensorrt --extra-index-url https://pypi.nvidia.com

If the node imports fine but then seems to hang forever on first run, give it real time before assuming it's broken (see above) - and if it's still failing afterward, pip install -r requirements.txt --force-reinstall followed by a restart is a fix that's worked for people in practice, not just theory. This is a general ComfyUI custom-node landmine, not specific to this pack: every node installs its Python deps into the same shared environment, so a half-finished install or a version conflict from something else you added can leave a node "installed" but non-functional.