English | 中文

ComfyUI custom nodes for rednote-hilab/dots.tts.

• Added an opt-in compile toggle to the bottom of Dots TTS Load Model using native PyTorch Inductor/Triton compilation.
• Added CUDA, Triton, Inductor, compile-length, and cudaMallocAsync compatibility guards. CUDA Graph Trees are disabled automatically with cudaMallocAsync while Triton compilation remains active.
• Compile works with SDPA and Flash Attention. Changing compile, model, device, dtype, or attention fully unloads the previous bundle before reloading.
• Fixed streaming vocoder LSTM compilation without enabling global Dynamo settings that could affect other ComfyUI nodes.
• Compiled graphs and static generation workspaces are cleared during manual unload.
• The terminal now displays Preparing/compiling until the first audio patch is ready. The first run for each length bucket is slower while PyTorch compiles it.

• Dots TTS Load Model
• Dots TTS Generate
• Dots TTS Voice Clone
• Dots TTS Whisper Transcribe

The loader catalog shows the official Rednote checkpoints first, then the drbaph BF16 conversions:

1. dots.tts Base FP32 (auto-download) - rednote-hilab/dots.tts-base
2. dots.tts SOAR FP32 (auto-download) - rednote-hilab/dots.tts-soar
3. dots.tts MF FP32 (auto-download) - rednote-hilab/dots.tts-mf
4. dots.tts Base BF16 (auto-download) - drbaph/dots.tts-base-bf16
5. dots.tts SOAR BF16 (auto-download) - drbaph/dots.tts-soar-bf16
6. dots.tts MF BF16 (auto-download) - drbaph/dots.tts-mf-bf16

• Quality first → dots.tts-soar
• Speed first → dots.tts-mf
• Training / fine-tuning → dots.tts-base

Downloaded model files are placed like this:

ComfyUI/
└── models/
    ├── dotstts/
    │   ├── common/
    │   │   ├── speaker_encoder.safetensors
    │   │   └── vocoder.safetensors
    │   ├── dots.tts-base/
    │   │   └── model.safetensors
    │   ├── dots.tts-soar/
    │   │   └── model.safetensors
    │   ├── dots.tts-mf/
    │   │   └── model.safetensors
    │   ├── dots.tts-base-bf16/
    │   │   └── dots.tts-base-bf16.safetensors
    │   ├── dots.tts-soar-bf16/
    │   │   └── dots.tts-soar-bf16.safetensors
    │   └── dots.tts-mf-bf16/
    │       └── dots.tts-mf-bf16.safetensors
    └── audio_encoders/
        ├── openai_whisper-large-v3-turbo/
        ├── openai_whisper-large-v3/
        ├── openai_whisper-medium/
        ├── openai_whisper-small/
        └── openai_whisper-tiny/

Small tokenizer/config assets are bundled in this custom node and separated by source model:

ComfyUI/
└── custom_nodes/
    └── Dots-TTS-ComfyUI/
        └── assets/
            ├── dots.tts-base/
            │   ├── added_tokens.json
            │   ├── chat_template.jinja
            │   ├── config.json
            │   ├── latent_stats.pt
            │   ├── llm_config.json
            │   ├── merges.txt
            │   ├── special_tokens_map.json
            │   ├── tokenizer.json
            │   ├── tokenizer_config.json
            │   └── vocab.json
            ├── dots.tts-soar/
            │   └── same small-file set
            └── dots.tts-mf/
                └── same small-file set

BF16 entries use the matching source-model assets. For example, drbaph/dots.tts-base-bf16 uses assets/dots.tts-base/.

Shared heavy assets come from drbaph/dots.tts-common and are stored under ComfyUI/models/dotstts/common/. The common repo files live at the repo root:

drbaph/dots.tts-common/speaker_encoder.safetensors
drbaph/dots.tts-common/vocoder.safetensors

At load time the node assembles an upstream-compatible runtime cache under runtime/ using links/copies from node assets, shared heavy assets, and the selected model weight. The loader uses Hugging Face directly and does not use HF mirrors.

max_audio_patches on both Generate and Voice Clone is the maximum audio patch budget for that generation, not a text-token limit. The default is 500. With the bundled configs, one patch is about 0.32 seconds, so 500 is about 160 seconds of audio budget. The model can stop earlier when it reaches EOS; very long text can hit the cap and end early. Voice Clone prompt audio paired with reference_text also consumes part of this budget.

Generation uses a live tqdm terminal progress bar with percentage, elapsed time, estimated remaining time, and iteration speed. Since Dots TTS decides its final length by EOS during generation, the live total is the configured max_audio_patches ceiling; after a successful early stop, the completed bar is normalized to the actual emitted chunk count.

The loader's optional compile toggle uses upstream's native torch.compile path with PyTorch Inductor and Triton. It is CUDA-only, requires a working Triton installation, and is compatible with both SDPA and Flash Attention. When ComfyUI uses the cudaMallocAsync allocator, the node automatically disables incompatible CUDA Graph Trees while keeping Inductor/Triton compilation enabled. Compilation is lazy: the first generation for each max_audio_patches length bucket is slower while the graph is compiled, then later generations reuse it. Compiled mode supports up to 1024 audio patches. Changing the model, device, dtype, attention, or compile setting fully unloads the active bundle before loading the new one; manual unload also clears compiled graphs and generation workspaces.

For the fastest model path, use the MF BF16 checkpoint with steps=4. Smaller max_audio_patches values can also select a smaller compile bucket and reduce compile time and workspace memory. Upstream recommends splitting long text into shorter segments and keeping voice-clone reference audio around 10 seconds.

Officially benchmarked: 24 languages — Chinese, English, Cantonese, Japanese, Korean, Arabic, Spanish, Turkish, Indonesian, Portuguese, French, Italian, Dutch, Vietnamese, German, Russian, Ukrainian, Thai, Polish, Romanian, Greek, Czech, Finnish, and Hindi. It may be able to do more languages but those are the ones officially benchmarked. Not all languages produce high quality results — you may need to experiment for yourself to see.

The language dropdown is kept to those 24 languages, plus auto and none: AR, YUE, ZH, CS, NL, EN, FI, FR, DE, EL, HI, ID, IT, JA, KO, PL, PT, RO, RU, ES, TH, TR, UK, VI.

ComfyUI-Manager (recommended): Open ComfyUI-Manager, search for Dots TTS, and click Install. ComfyUI-Manager will handle everything automatically.

Manual helper install with uv:

python -m uv pip install -r requirements.txt

Manual helper install with pip:

python -m pip install -r requirements.txt

The installer protects ComfyUI's core runtime packages and will not automatically upgrade torch, torchaudio, torchvision, transformers, or pydantic.

Dots upstream recommends recent transformers and pydantic v2. This node warns about those versions instead of changing them automatically, because surprise upgrades can break other ComfyUI nodes.

Audio file I/O uses soundfile first. Dots' speaker feature path has a torchaudio-free fallback for broken torchaudio installs, though the original torchaudio/Kaldi fbank path is used when available.

• Dots TTS upstream repository
• Official Dots TTS Base model
• Official Dots TTS SOAR model
• Official Dots TTS MF model
• drbaph Dots TTS Base BF16 model
• drbaph Dots TTS SOAR BF16 model
• drbaph Dots TTS MF BF16 model
• drbaph Dots TTS shared common files
• OpenAI Whisper large-v3-turbo

@article{dotstts2026,
  title   = {dots.tts Technical Report},
  author  = {dots.tts Team},
  journal = {arXiv preprint},
  year    = {2026},
}

Released under Apache-2.0.