Load Rife Tensorrt Model

This node looks like a boring dropdown with two options, but it's doing the heavy lifting for the whole pack. It loads a RIFE frame-interpolation model as a TensorRT engine - and if that engine doesn't exist yet, it downloads the model, compiles the engine, and only then hands it over. First run is slow. Every run after that is instant. That asymmetry is the entire thing to understand here.

RIFE (Real-Time Intermediate Flow Estimation) is the classic video frame interpolator: give it two frames, it estimates the motion between them and synthesizes what happened in between. That's what powers slow-mo and frame-rate smoothing in a dozen ComfyUI packs. The plain-PyTorch version works fine but crawls on long clips. This pack compiles the same model into an NVIDIA TensorRT engine, which is the "add pixels over time, fast" rung of the upscaling ladder - frame interpolation is the companion to video upscalers, not a substitute for them. It adds motion between frames, it does not invent detail.

The two inputs that matter

• model - rife49_ensemble_True_scale_1_sim by default, with rife47 and rife48 as the alternatives. Ignore the suffix noise; the version number is what matters. rife47 is the fastest, rife48 the middle ground, rife49 the latest and most accurate (and the default for a reason).
• precision - fp16 or fp32. fp16 is the default and the one you want for speed. fp32 buys a little accuracy at the cost of a slower engine. Unless you're chasing artifacts on specific footage, leave it on fp16.

Its single output, rife_trt_model, is a RIFE_TRT_MODEL handle - not an image, not a tensor. It wires straight into the rife_trt_model input of the pack's other node, ⚡ Rife Tensorrt. That's the whole two-node flow: load here, interpolate there.

What actually happens under the hood

The node looks for a cached engine in ComfyUI/models/tensorrt/rife/. The filename encodes everything - model, precision, the 256→3840 resolution profile, and the installed TensorRT version - so each combination is its own engine file. If the engine is missing, it checks ComfyUI/models/onnx/ for the .onnx; if that's missing too, it downloads it from the author's HuggingFace repo (yuvraj108c/rife-onnx). Then it builds the engine in-process with polygraphy - no trtexec binary to hunt down. On a decent GPU, figure a few minutes the first time; the console will print "Building TensorRT engine" so you know it's alive. Because the engine file embeds the TRT version, upgrading TensorRT silently forces a rebuild. That's by design, not a bug.

Install

ComfyUI Manager: search "ComfyUI Rife TensorRT" and install. Or by hand:

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

Dependencies are light: einops, colored, polygraphy, and tensorrt. No model files to download manually - that's the 2026 rewrite's whole point. One heads-up: requirements.txt leaves TensorRT unpinned while the (older) pyproject.toml pins tensorrt==10.4.0; the author tests against TRT 10.14 / CUDA 13 / Python 3.12, so don't be shocked if an old wheel installs. And yes, this pack is non-commercial (CC BY-NC-SA 4.0) - fine for personal slow-mo, a problem if this is for client work.

The classic complaint, mostly from Windows users pre-2026, was engine building failing without a CUDA toolkit. The author ripped out the cuda-python dependency specifically to kill that install pain. If you're on an old checkout, update before you blame your GPU.