Qwen3-TTS Finetune

Qwen3-TTS Finetune is the serious node in this pack, and the reason most people install it: it trains a dedicated model on your own voice data. Where Voice Clone extracts a voice from a short reference clip, fine-tuning teaches the whole model to be that voice - the difference between a thin imitation and a stable, high-fidelity speaker. It's a genuine training loop running inside ComfyUI, with all the VRAM-management muscle that implies, and it's the closest thing this pack has to LoRA training, except it's a full fine-tune of the Base model rather than a small adapter.

Fair warning before you dive in: this is not a two-minute node. It expects a properly prepared pipeline (Dataset Maker → Data Prep → this node), downloads the Base model if you haven't got it, and a real fine-tune on a 1.7B model takes a while on a consumer GPU. If you just want a character voice for a quick clip, Voice Clone is the cheaper path. If you need a speaker you can trust across hours of content, fine-tune.

What it trains

The key input is init_model, and it's restricted to the Base variants (Qwen/Qwen3-TTS-12Hz-1.7B-Base or the 0.6B) - the CustomVoice/VoiceDesign checkpoints are excluded because you can't sensibly train on top of a preset-voice model. The node loads the Base model, enables gradients on everything, and trains with a proper optimizer loop: AdamW (or 8-bit AdamW if bitsandbytes is installed and use_8bit_optimizer is on), bf16 mixed precision by default (auto-falling back to fp32 on pre-Ampere GPUs), and the training data comes from the train_jsonl you wired in.

Two design choices the author got right, and they're baked into the tooltips:

• lr defaults to 2e-6, not the 2e-5 Qwen ships with. The tooltip explains: "Qwen default (2e-5) is too aggressive for small batches, causing noise output." If you've ever fine-tuned a TTS and gotten crackly garbage, this is that exact failure pre-solved. Leave it at 2e-6.
• speaker_name is how you summon the trained voice later. The node outputs it back to you (custom_speaker_name), and you type it into Qwen3-TTS Custom Voice's custom_speaker_name field to generate with the trained voice - it won't show up in the preset dropdown.

The inputs that matter

The VRAM trinity is where beginners should look first:

• gradient_checkpointing (default on) - recomputes activations instead of storing them, ~30-40% VRAM savings.
• use_8bit_optimizer (default on) - 8-bit AdamW, ~50% less optimizer VRAM. Requires bitsandbytes; the node silently falls back to regular AdamW and prints a note if it's missing.
• batch_size (default 2) + gradient_accumulation (default 4) - effective batch size is batch_size × gradient_accumulation. Keep batch_size low and let accumulation do the work on a small GPU.

save_every_epochs (default 1) writes per-epoch checkpoints into output_dir; resume_training picks up the most recent one and continues. save_optimizer_state doubles checkpoint size for a perfect resume - leave it off unless you're doing long multi-session trains. Both outputs, model_path and custom_speaker_name, are strings you feed back into the loader and generator.

Installing & troubleshooting

Pack install as usual: ComfyUI Manager (search "ComfyUI-Qwen3-TTS") or git clone https://github.com/DarioFT/ComfyUI-Qwen3-TTS into custom_nodes, pip install -r requirements.txt by hand. The qwen-tts transformers==4.57.3 pin is the sharpest edge here because you're about to add accelerate, librosa, tensorboard and a training stack to a shared environment - the README's advice about a separate Python env is aimed squarely at this node.

The classic failures: OOM during training (drop batch_size, keep gradient checkpointing on), noisy/crackly output (your learning rate drifted from 2e-6, or your dataset is bad - clean single-speaker clips), and a GPU pinned at 100% that never finishes (the known Qwen3-TTS hang; it's rarer in training but the kill-and-restart advice still applies). Training progress goes to the ComfyUI UI, so "is it alive?" is answered by watching the console.

When it works, load the checkpoint through Qwen3-TTS Loader's local_model_path (point at output_dir/epoch_X) and generate with Custom Voice using your speaker_name. That's the whole arc: data → train → evaluate → use.