ParlerTTS_LoadModel

The whole ComfyUI_ParlerTTS pack is two nodes, and this is the boring half - one text field, no sliders, no toggles. Don't let that fool you. ParlerTTS_LoadModel is where most of the pack's pain lives, because it's the node that downloads the model. Once you've sat through a first run, everything after is smooth.

Parler-TTS is Hugging Face's lightweight text-to-speech model (built with Stability AI) that turns a script and a written description of the speaker into natural-sounding audio. This loader grabs the checkpoint from Hugging Face and hands the whole thing to the Sampler node. If you've got an EchoMimic or talking-head workflow and you want the voice to come from inside ComfyUI instead of a separate Python script, this is the pack that gets you there.

How it works

Mechanically it's a thin wrapper around Hugging Face's own code. The node calls snapshot_download to pull the repo into your HF cache, then loads a ParlerTTSForConditionalGeneration model plus its tokenizer and bundles them into a dict it calls a PTTSMODEL. Nothing runs yet - you're not generating audio here, just loading weights. That's the whole job.

There's one detail worth knowing: if your repo_id contains "large", the loader uses trust_remote_code=True because the big checkpoint carries its own modeling code. The mini models don't need it. You don't have to do anything about this - it just explains why a "large" download can behave differently.

The one input

• repo_id - which checkpoint to load. The default is parler-tts/parler_tts_mini_v0.1, and if you leave the field empty it silently downloads that. Otherwise you can type any of the family:
  • parler-tts/parler-tts-mini-v1 (~300M params, the daily-driver)
  • parler-tts/parler-tts-large-v1 (~1.6B, better but heavier)
  • parler-tts/parler-tts-mini-jenny-30H - 30 hours of one specific voice, and it requires the keyword "Jenny" in your voice description later
  • any local absolute path (X:/models/parler-tts-mini-v1), which is how you run fully offline

Output is a single model socket (PTTSMODEL), which is the only thing the ParlerTTS_Sampler accepts. It's a pack-internal type - don't try to wire it into anything else.

Installing it

Same story as every smthemex pack. In ComfyUI Manager, search for ComfyUI_ParlerTTS and install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/smthemex/ComfyUI_ParlerTTS.git
pip install -r requirements.txt

Restart ComfyUI and you're done. The dependency list is unusually light for a TTS pack - transformers and torch - because the audio codec (DAC) is vendored inside the repo. That's a real improvement: older versions needed descript-audio-codec installed, and that library is notorious for fighting everything else in your environment. Community reports from late 2024 called that out specifically; v1.1 got rid of it.

Where people get burned

First load downloads the model, and a mini checkpoint is roughly a gigabyte-plus from Hugging Face. No network, no load - that's the #1 "why is this stuck" moment, and it's not stuck, it's downloading.

The README also flags a transformers version sensitivity: versions newer than ~4.53.0 have caused errors, with a patch shipped July 2025. If your ComfyUI Manager auto-updates transformers and this node suddenly breaks, that's your suspect. And one genuine trap from the ecosystem at large: custom node packs install into the same Python environment with no isolation, so a bare pip install -r requirements.txt can clobber versions another node depends on. This pack's deps are tame enough that it rarely matters - but it's why people run Manager's own install instead of hand-rolling it.