Spark_TTS_Creation

You've got the images, you've got the video - the part that always stalls is the voiceover. Spark_TTS_Creation is the node that fixes that, by dropping a real text-to-speech model into your ComfyUI graph. Type a line, pick a gender, pitch and speed, and out comes a 16 kHz waveform you can save, play, or feed straight into a video render. No API key, no account, no phoning home to a cloud service - it runs entirely on your machine.

It's one half of the ComfyUI_Spark_TTS pack, the "voice from nothing" half. The sibling node, Spark_TTS_Clone, handles cloning a specific voice; this one generates a clean, controllable speaker. Under the hood it's Spark-TTS 0.5B from the SparkAudio team, an LLM-based TTS model. The "0.5B" is the part that matters for you: it's the small, fast end of the TTS spectrum, the kind of model that loads quickly, runs on modest GPUs, and still sounds a world better than the robot voices you grew up with.

How it works

Mechanically, Spark-TTS is an autoregressive LLM that turns text into semantic speech tokens, then a biCodec vocoder converts those tokens into an actual waveform. The node vendors the official Spark-TTS library - there's a bundled sparktts folder - and drives it through a thin core wrapper. On the first run it loads weights from ComfyUI/models/TTS/Spark-TTS/Spark-TTS-0.5B/; if that folder is empty, it tries to auto-download the model before you even queue up. The output is a standard ComfyUI AUDIO dict (waveform plus sample rate), so it plugs into the same audio save/preview nodes you already use for any other AUDIO.

The inputs that matter

Only a few of these are worth touching on day one:

• text - the thing the voice actually says.
• gender (female/male) plus pitch and speed, each from very_low to very_high. This is the real control this node sells: dialing in a voice instead of rolling the dice.
• max_new_tokens (default 2020) - raise it if a clip gets cut off mid-sentence.
• keep_model_loaded (True) - right call for most people. The 0.5B model is small, so keeping it warm costs little VRAM and makes every rerun faster.
• use_cpu (False) - the escape hatch for GPUs ComfyUI can't schedule (some AMD cards); it also routes to MPS on Macs.

The sampling quartet (temperature 0.7, top_k 30, top_p 0.7) behaves like samplers-as-usual - leave them alone unless output starts sounding flat or chaotic.

Outputs

Two outputs: Audio, the thing you actually keep, and Node Status - a string that mostly exists because this pack would rather hand you a readable error in the graph than silently die. If the audio comes back as silence, that string is the first place to look.

Installing it

Easiest route:

# ComfyUI Manager: search "Spark-TTS" → Install → restart

Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/KERRY-YUAN/ComfyUI_Spark_TTS
cd ComfyUI_Spark_TTS
# use your ComfyUI Python (e.g. python_embeded/python.exe on Windows):
pip install -r requirements.txt

The requirements are mostly light (transformers, soundfile, einops, omegaconf), but the list also demands torch and torchaudio - which you already have. The README's warning about keeping those compatible with your existing PyTorch install is the one to respect. The model auto-downloads on first use; the manual fallback is running Model_Download.bat in the node folder.

Where people get burned

The failure you'll actually hit: the model folder isn't where the node expects it, and Node Status fills with "Please run Model_Download.bat or check paths." The auto-downloader is one-shot per session and needs internet plus Git on your PATH (the companion preset repo is a git clone), so when it silently fails, running Model_Download.bat is the fix. A "no semantic tokens found" error usually means max_new_tokens is too low for your text. And on CPU, expect slow - that's not a bug, that's an autoregressive model grinding token by token.

Honest take: this is a fast, capable-enough narrator, not an emotional ElevenLabs impersonator. If you need clone-perfect voices, the other node in this pack - or the rival TTS packs the community argues over - will take you further.