Qwen3 ForcedAligner Loader

This loader is even simpler than the ASR one - it's the "I only exist so the aligner node has a model" node. If you don't need timestamps, you don't need this at all. If you do - subtitles, karaoke, lip-sync prep, dataset labeling - it's the mandatory first step, and the only way to get it wrong is to forget it's eating ~4GB of VRAM.

What it does

One input, one output. model_name is a dropdown with exactly one entry - Qwen3-ForcedAligner-0.6B - so there's no real choice to make; it's just a model identifier. The output, aligner (type QWEN3_ALIGNER), wires only into the pack's Qwen3 Forced Align node. It loads a separate ~0.6B model whose whole job is matching the words of a known transcript to their positions in an audio file - something the transcription model itself doesn't give you.

How it works

Same playbook as its ASR sibling: on first use it downloads Qwen/Qwen3-ForcedAligner-0.6B from HuggingFace into ComfyUI/models/Qwen3-ASR/Qwen3-ForcedAligner-0.6B/, then loads it in bfloat16 on cuda:0. After that it's a local disk read, and the model sits in VRAM until you restart or purge it.

Because the aligner is a separate model from the ASR model, you pay twice: a second multi-gigabyte download and a second VRAM footprint. The README puts it at ~4GB VRAM, so on an 8GB card, running the 1.7B ASR model and the aligner in the same graph is already over budget before your video pipeline adds anything.

Installing

The full pack:

cd ComfyUI/custom_nodes
git clone https://github.com/shumoLR/Comfyui_SynVow_Qwen3ASR
cd Comfyui_SynVow_Qwen3ASR
pip install -r requirements.txt

Or via ComfyUI Manager: search "Comfyui_SynVow_Qwen3ASR", install, restart. Run this loader once ahead of time so the model download happens while you're not mid-workflow.

Common issues

• CUDA or out-of-memory at load. Same GPU-only story as the ASR loader - bf16, cuda:0, no fallback. On an 8GB card, drop the ASR model to 0.6B when you need alignment in the same session.
• Model seems to download every run. That's the empty-folder check before the download finishes; once it's complete, it's a local load.
• The aligner output won't plug in anywhere. It only mates with Qwen3 Forced Align - there's nothing else in the pack that accepts it.

If you only need plain transcripts, skip this node entirely and save yourself the download and the VRAM.