// the find
comfyanonymous/ComfyUI_TensorRT
A ComfyUI custom node that converts Stable Diffusion checkpoints into NVIDIA TensorRT engines, giving you measurably faster inference on RTX GPUs. The tradeoff is upfront engine compilation time (3–60 minutes depending on model) and engines that are locked to a specific GPU architecture. For anyone running ComfyUI on an RTX card and doing high-volume generation, this is worth knowing about.
Engine compilation happens once and then gets reused across sessions, so the upfront cost amortizes quickly if you run hundreds of generations. Dynamic engines are a sensible design — you specify min/opt/max for resolution and batch size, and the engine stays flexible without requiring a recompile for every resolution. The workflow JSON files are a genuine convenience; you can load them directly into ComfyUI without guessing what nodes to wire together. Coverage of the major model families (SD1.5, SDXL, SVD, Flux, AuraFlow) means it's not just an SD1.5 trick.
No ControlNet or LoRA support, which rules out a significant chunk of practical ComfyUI workflows — this isn't a footnote, it's a hard blocker for anyone whose pipelines depend on them. The VRAM floor is steep: Flux requires 24GB+, SVD-XT requires 24GB+, so most consumer cards are excluded from the more interesting models. Last push was October 2024 and the repo is essentially a thin wrapper (~two Python files) with no tests, so when TensorRT or ComfyUI internals change, there's not much here to catch the breakage. Engines are GPU-architecture-specific, meaning a compiled engine from one machine is useless on a different card — something the README buries.