ComfyUI-EMALightning

EMA Lightning for ComfyUI

Turkish text to speech in ComfyUI with EMA Lightning (canberkkkkkk/ema-lightning, code canberk7/ema-lightning): an 8.6M-parameter flow-matching model and vocoder (34 MB) that speaks 48 kHz Turkish, reads numbers, dates, times and amounts aloud through a Turkish text normalizer, and reports when every word is spoken. It runs inside the ComfyUI process through the official ema-lightning package, on the GPU or the CPU; the nodes never download anything.

Unofficial. This node pack is an independent integration. It is not affiliated with or endorsed by the EMA Lightning authors or by Comfy Org.

日本語の概要は README.ja.md にあります。

The shipped tts_basic workflow through the real ComfyUI queue: the saved audio and the word times the node returned.

What it does

| Node | Purpose | | --- | --- | | EMA Lightning Loader | Model folder from models/ema_lightning, device (GPU or CPU), the scheduler's batch size, and the optional compiled "lightning" mode | | EMA Lightning Text to Speech | Text, seed, speed (0.25 to 4), sample rate (48 / 24 / 16 / 8 kHz) → AUDIO, the word times (JSON) and a report |

Connect the audio to ComfyUI's Save Audio (Advanced) (FLAC, MP3 or Opus) or Preview Audio. Two workflows are included: tts_basic (FLAC) and tts_announcement_mp3 (an announcement with a time and an amount, saved as MP3).

How far it was checked

Windows 11, RTX 5090, ComfyUI v0.38.0, torch 2.14.1, ema-lightning 1.1.2 (details in docs/VERIFICATION.md):

• Bit-identical to the official package on the GPU. The documented usage (EMA(device).say(text, speed=..., seed=..., sample_rate=...)) in a separate environment gives exactly the same waveform and word times as the node in 9 of 9 cases: short and long texts, numbers and dates, three seeds, speeds 0.75 / 1.0 / 1.5, 48 / 24 / 16 kHz, and batch sizes auto, 1 and 8.
• On the CPU the node equals the official package run under the same process conditions (9 of 9). Between processes, CPU results can differ by up to about 6e-7 (-124 dB) depending on what else the process imported (seen with ComfyUI's own model management loaded); the same holds for the official package alone.
• Real ComfyUI queue: both shipped workflows run; the saved 16-bit FLAC equals the official float audio rounded to 16 bits (largest difference half a 16-bit step).
• Clean install: fresh ComfyUI v0.38.0, fresh Python 3.12 environment, git clone, pip install -r requirements.txt, the model downloaded with tools/download_model.py (SHA-256 verified): both workflows give the same audio samples and word times as the development environment.
• CI (GitHub Actions; Ubuntu and Windows; Python 3.11 / torch 2.8.0 and Python 3.12 / torch 2.14.1): the tests and a real ComfyUI queue on the CPU with the real model.

Not checked: Linux or macOS with a GPU, AMD / Apple GPUs, the lightning mode with a working torch.compile (Triton), and speech quality against the model card's benchmarks (the pack reproduces the official audio; it does not re-evaluate the model).

Install

1. ComfyUI with the V3 node API (v0.38.0 was used), Python 3.11+ (required by ema-lightning and normalizer-tr). Any NVIDIA GPU or just the CPU (the model is 34 MB).

2. Clone this repository into ComfyUI/custom_nodes/ and install its requirement (the official package, which brings the Turkish normalizer):

   git clone https://github.com/hiroki-abe-58/ComfyUI-EMALightning custom_nodes/ComfyUI-EMALightning
   python -m pip install -r custom_nodes/ComfyUI-EMALightning/requirements.txt
   

3. Download the model (about 34 MB). The script pins the revision this pack was checked with and verifies every file's SHA-256:

   python custom_nodes/ComfyUI-EMALightning/tools/download_model.py --models-dir models/ema_lightning
   

   This creates models/ema_lightning/ema-lightning/. Another location works through extra_model_paths.yaml with the key ema_lightning.

4. Restart ComfyUI and open workflows/gui/tts_basic.json.

Using the nodes

• Text is Turkish. Numbers, dates (15 Ekim 2026), times (14:30), amounts (1.250.000 TL) and symbols are read out by the package's normalizer; the word times list the words as spoken ("14:30" comes back as words). Text is lowercased with Turkish rules, so for example "ComfyUI" is listed as "comfyuı". Long texts are split into sentences.
• Seed: the same text, seed, settings and device give the same audio. CPU and GPU give different audio for the same seed.
• batch_size (Loader): how many sentences or audio windows the package's scheduler runs at once. auto is the package's own default: it measures the throughput once per device and caches the result in its cache folder (~/.cache/ema_lightning, or under XDG_CACHE_HOME). The audio can differ very slightly between batch sizes (largest seen: 9e-4) when a text needs more audio windows than the batch holds; a batch of 1 differs from all larger ones. The measured value is not stable: on the same RTX 5090 one measurement gave 128 and another (fresh install) 8. For audio that repeats exactly across installs and machines, choose a fixed batch size.
• lightning (Loader, GPU only, off by default): the package's fast mode, recorded once per session. On this Windows machine torch.compile had no Triton, so the package recorded CUDA graphs without compiling: 190 graphs in 158 s, then a 7.9 s announcement in 12 ms. The audio differed from the plain path by up to 6.8e-4 (the package accepts up to 1e-2), so it is not bit-identical. The package also switches on cuDNN benchmark mode for the whole process when lightning starts, which affects other models in the same ComfyUI session.
• Word times (JSON: text, start, end in seconds) are useful for subtitles or lip-sync timing.
• The model stays loaded for the session (the package's scheduler thread keeps it alive; it is 34 MB).

Measured on this setup

RTX 5090, batch size auto (128): a 7.9 s announcement in 0.036 s (about 220 times faster than real time); a short sentence in about 0.05 s after the first call. CPU (Intel Core Ultra 9 285K): a 5.7 s sentence through the whole ComfyUI queue in about 1 s.

Licenses

• This node pack: MIT (LICENSE). It contains no third-party code.
• ema-lightning and normalizer-tr (installed packages) and the model (not included; downloaded only by tools/download_model.py): Apache-2.0. The model card adds that whoever deploys it is responsible for using it lawfully, including disclosure and consent rules for synthetic speech. See NOTICE.

Development

python -m pip install pytest ruff
python tools/download_model.py --models-dir refs/ema_lightning
EMA_COMFYUI_DIR=<ComfyUI checkout> EMA_MODEL_DIR=refs/ema_lightning/ema-lightning python -m pytest tests
python scripts/ci_queue_check.py --comfyui <ComfyUI checkout> --model refs/ema_lightning/ema-lightning
python scripts/run_parity.py --comfyui <ComfyUI checkout> --model refs/ema_lightning/ema-lightning --out pack.npz