What is this?
A ComfyUI custom node that runs the MiniMax-H3 VAE as a compiled TensorRT engine instead of the regular one, making it up to 1.7x faster.
A ComfyUI custom node that runs the MiniMax-H3 VAE as a compiled TensorRT engine instead of the regular one, making it up to 1.7x faster.So it's not a separate app — it lives inside ComfyUI?Right — it's a ComfyUI custom node, not a standalone program. After installing, you get new VAE nodes inside your ComfyUI workflow.
Right — it's a ComfyUI custom node, not a standalone program. After installing, you get new VAE nodes inside your ComfyUI workflow.How do I get it running?Install the node like any ComfyUI extension, download the encoder and decoder .onnx files (plus .data files if present), and put them into ComfyUI/models/vae.
Install the node like any ComfyUI extension, download the encoder and decoder .onnx files (plus .data files if present), and put them into ComfyUI/models/vae.Is it ready to use right after installing?Not quite — you first compile a TensorRT engine from the ONNX with the MiniMax-H3 TRT VAE Compiler node (or the compile.py script outside ComfyUI). After that, the Loader node just loads the finished engines.
Not quite — you first compile a TensorRT engine from the ONNX with the MiniMax-H3 TRT VAE Compiler node (or the compile.py script outside ComfyUI). After that, the Loader node just loads the finished engines.Can I run it on a Mac, or is it NVIDIA-only?TensorRT only runs on NVIDIA GPUs, so this is for Windows or Linux machines with NVIDIA cards; macOS is not a target.
TensorRT only runs on NVIDIA GPUs, so this is for Windows or Linux machines with NVIDIA cards; macOS is not a target.Will it run on my GPU?The default decoder wants 12GB+ of VRAM. If your card has less, use the quantized w4a16awq decoder instead.
The default decoder wants 12GB+ of VRAM. If your card has less, use the quantized w4a16awq decoder instead.What exactly is MiniMax-H3?It's the AI model from MiniMax-AI whose VAE this node accelerates. The project doesn't touch the model itself — it only repackages the VAE as ONNX and runs it through TensorRT inside ComfyUI.
It's the AI model from MiniMax-AI whose VAE this node accelerates. The project doesn't touch the model itself — it only repackages the VAE as ONNX and runs it through TensorRT inside ComfyUI.Can I use it with VAEs from other models?No — it's built specifically for the MiniMax-H3 VAE. You download that model's encoder and decoder ONNX files and compile TensorRT engines from those.
No — it's built specifically for the MiniMax-H3 VAE. You download that model's encoder and decoder ONNX files and compile TensorRT engines from those.Is this an official MiniMax tool?No — it's a community project by developer lihaoyun6. The README just credits ComfyUI and MiniMax-AI's MiniMax-H3 as the underlying projects.
No — it's a community project by developer lihaoyun6. The README just credits ComfyUI and MiniMax-AI's MiniMax-H3 as the underlying projects.What are similar projects?FreeVideo, server, ComfyUI-Ref2VA-VSA, InstinctFlash, TensorRT, ComfyUI-YCNodes-MiniMax-H3, Hyperloom, onnxruntime
FreeVideo, server, ComfyUI-Ref2VA-VSA, InstinctFlash, TensorRT, ComfyUI-YCNodes-MiniMax-H3, Hyperloom, onnxruntimeWhich topics is it in?inference-optimization, comfyui-nodes, tensorrt, inference-engine, comfyui, optimization, onnx, quantization, video-generation, inference, nvidia, performance
inference-optimization, comfyui-nodes, tensorrt, inference-engine, comfyui, optimization, onnx, quantization, video-generation, inference, nvidia, performance