(Down)load RIFE TensorRT Model

The name tells you the whole pitch: this node loads a RIFE model, and if it can't find one, it automatically downloads and builds it. That's the entire point of the "Auto" in the pack name - and it's aimed squarely at the older RIFE-TensorRT packs that made you hand-build .engine files with export scripts before you could interpolate a single frame. People genuinely got stuck on that step for hours. This removes it.

AutoLoadRifeTensorrtModel is the loader half of the pack. You run it once, it hands you a RIFE_TRT_MODEL, and you wire that into AutoRifeTensorrt which does the actual interpolation.

How it works

The first time you run it, two things happen:

1. It checks ComfyUI/models/onnx/<model>.onnx. Missing? It downloads it from HuggingFace - primary source is huchukato/favs, with yuvraj108c/rife-onnx and marduk191/rife as automatic fallbacks if the first mirror fails.
2. It checks ComfyUI/models/tensorrt/rife/ for a compiled engine matching your model, precision, and resolution profile. No match? It compiles the ONNX into a TensorRT engine with polygraphy on the spot.

That engine build is the part to respect: it takes minutes and eats VRAM, and the first time you run the node it looks exactly like ComfyUI froze. It hasn't. The resulting .trt file is cached forever, so the pain is a one-time deal. The engine filename even encodes everything about it (model, precision, profile, batch, TRT version), so changing any setting rebuilds a fresh engine rather than corrupting the old one.

The inputs that matter

• model - three tested RIFE variants: rife49_ensemble_True_scale_1_sim (default, most accurate), rife48... (the balance pick), rife47... (fastest). For frame interpolation, rife49 is the sensible default; the quality gap to 47 is small but real at high multipliers.

• precision - fp16 (default) or fp32. fp16 is roughly half the VRAM and nearly indistinguishable in output for interpolation. Reach for fp32 only if you're chasing artifacts nobody else can see. (The usual fp16-vs-fp32 story, and here the fast side is the right side.)

• resolution_profile - this is the one that bites people. The TensorRT engine is compiled for a resolution range:

  • small (384–1080px) - default, least VRAM
  • medium (672–1312px)
  • large (720–1920px) - for 4K-ish content
  • custom - uses a CustomResolutionConfig node you connect below

  Rule of thumb: pick the smallest profile that covers your output resolution. A smaller range means a smaller, faster-to-build engine and less VRAM. The README's rough budgets: ~2GB for small, ~4GB medium, ~6GB large.

• custom_config (optional) - a RIFE_RESOLUTION_CONFIG from RIFE Custom Resolution Config, only consulted when profile is custom.

The single output, rife_trt_model, plugs into AutoRifeTensorrt's rife_trt_model input.

Installing it

ComfyUI Manager: search ComfyUI-RIFE-TensorRT-Auto. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/huchukato/ComfyUI-RIFE-TensorRT-Auto.git

Restart ComfyUI. On first import the pack auto-installs its dependencies and the TensorRT wheels matching your CUDA (12 for RTX 30/40, 13 for RTX 50). The real prerequisite is the NVIDIA CUDA Toolkit installed on the system - the pack detects it via nvcc, CUDA_PATH, or CUDA_HOME. If auto-install fails, the fallback is python install.py in the pack folder, or pip install -r requirements_cu12.txt / requirements_cu13.txt by hand.

Troubleshooting

• "CUDA toolkit not found" warning - you need the actual CUDA Toolkit installed, not just the driver. Grab it from NVIDIA's CUDA download archive and set CUDA_PATH.
• First run takes 5+ minutes, fans spin up, nothing happens - engine build. Wait it out; subsequent runs are instant.
• Engine build crashes on small VRAM cards - drop to a smaller profile or fp16. If you're under ~2GB free, the small profile itself may be too much.
• Auto-install keeps failing - the pack writes marker files (.tensorrt_auto_installed / .tensorrt_auto_install_failed) in its folder. Delete the failed marker and restart to retry, or set DISABLE_TENSORRT_AUTO_INSTALL=true and install the wheels yourself.

If the model download fails from all three mirrors, the error tells you to run export_onnx.py and drop the file in models/onnx/ - a rare edge, but it's the documented escape hatch. And remember the pack is CC BY-NC-SA: great for your own projects, not for reselling the output in a product.