ORPH_Load

ORPH_Load is the front door of the ComfyUI-Orpheus pack. Everything else in the pack hangs off what this node outputs: the actual Orpheus 3B text-to-speech model, loaded and ready to talk. If you've ever loaded a checkpoint or a GGUF in ComfyUI, the shape will feel familiar - pick a file from a dropdown, get a model you can't see out the other end - except this model doesn't make pictures. It makes voices.

Orpheus, for context, is Canopy Labs' open-weight TTS model built on a Llama 3.2 3B backbone. It's not the lightest TTS on the block - Kokoro and Piper exist precisely because not everyone wants a 3B autoregressive model in their pipeline - but it's the one that gives you decent, expressive speech from a proper language model. And in ComfyUI, this pack is the way you actually run it.

How it works

Under the hood ORPH_Load does three things, and the clever part is the middle one. It reads a bundled config (orpheus-config.json) that describes a Llama-style 3B model - 28 layers, hidden size 3072, RoPE scaling for long context - then reads the safetensors file you point it at and checks the shape of lm_head.weight to sniff out the actual vocabulary size. That's how one loader handles both the base orpheus-3b and fine-tunes like orpheus-ft whose tokenizers are a hair different. It then hands the whole thing to Hugging Face's LlamaForCausalLM and returns it as an ORPH_MODEL.

It also registers a custom model folder - ComfyUI/models/orpheus/ - the first time the pack loads, which is where the dropdown below gets its entries.

The inputs and outputs that matter

There's exactly one input: model, a dropdown of every .safetensors file sitting in models/orpheus. That's it. Nothing to tune, nothing to set.

The one output, ORPH_MODEL, wires straight into the model input of ORPH_Sample. That's the only place it goes.

Installing it

Two options, same result. In ComfyUI Manager, search for ComfyUI-Orpheus and install. Or do it by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/AustinMroz/ComfyUI-Orpheus

Restart ComfyUI after either. The pack itself has no requirements.txt - the SNAC audio codec is bundled inside the repo, and the other dependencies (torch, transformers, safetensors) already come with ComfyUI. The real install work is the weights:

• Orpheus 3B (fp16 safetensors, roughly 6 GB) from AustinMroz/orpheus-3b-f16-safetensors → ComfyUI/models/orpheus/

The README is blunt about where this lands in the lifecycle: "Heavily WIP." It works, but it's a hobbyist tool, not a polished release.

Common issues

• The dropdown is empty. The file isn't in models/orpheus/ - note that folder doesn't exist until the pack has loaded once, so create it by hand and drop the safetensors in, then restart.
• Every Orpheus node shows up broken at startup. The pack imports torchaudio at module load, and stock ComfyUI's pip install doesn't include torchaudio. If your env is missing it, pip install torchaudio (matching your installed torch version) fixes the whole pack at once.
• VRAM anxiety. The model is moved to GPU for generation and back to CPU afterwards, so it won't sit resident and eat your card between runs. Still, a 3B Llama is a chunky addition to a graph that's already running a diffusion model - budget for it.

ORPH_Load is boring on purpose. It's the "put the model in" step. The interesting stuff starts at ORPH_Sample.