Scenema Audio comes to ComfyUI
Scenema Audio, the expressive text-to-speech (TTS) model we announced in May, is now a native ComfyUI custom node package. Install it through ComfyUI Manager, load the workflow, and generate performance quality speech with the same model that powered the original announcement demos.
The model reads a prompt as direction, not just text. Laughs, whispers, voice cracks, breath, accents, and emotional shifts can land within a single generation, all cued in natural language.
What it sounds like
Here is a quick taste of the shipped presets. Each of these voices is a one-click starting point in the dropdown, ready to queue as-is or edit.
Install
Two ways to add it to your ComfyUI. Either way, the pre-wired workflow drops into your Workflows sidebar automatically. One click to load, everything already connected.
ComfyUI Registry (recommended). Open ComfyUI Manager, click Custom Nodes Manager, search “Scenema Audio”, click Install. Restart ComfyUI when it finishes. Nothing else to configure.
GitHub. Clone into your custom_nodes/ and install the requirements:
cd custom_nodesgit clone https://github.com/ScenemaAI/ComfyUI-ScenemaAudio.gitpip install -r ComfyUI-ScenemaAudio/requirements.txtRestart ComfyUI when it finishes.
All model weights auto-download from HuggingFace on first use. Total download is around 30 GB, one time.
HuggingFace token
The text encoder is google/gemma-3-12b-it, which is gated. Before you run your first generation:
- Visit https://huggingface.co/google/gemma-3-12b-it and click “Agree and access repository”.
- Create a token at https://huggingface.co/settings/tokens (read scope is fine).
- Register the token with either
huggingface-cli loginorexport HF_TOKEN=hf_...before launching ComfyUI.
If you skip this, the first generation fails with clear instructions.
The workflow
After install, open ComfyUI’s Workflows sidebar (left edge). You’ll see a Scenema Audio folder with the workflow inside. Click to load — every node is already wired.

Six nodes, three logical sections.
- Loaders on the left. Model Loader and VAE Loader, shared across everything.
- Voice cloning at the bottom left. Optional. Upload a reference audio file, feed it through Encode Reference, connect the resulting latent to Generate’s
ref_latentinput. - Generate in the center. The single node that does all the work. All the writing happens here.
- Preview Audio on the far right, so you can hear the result inline.
Preset voices
Every voice from the original announcement demos is a one click preset. Pick one, all the fields fill in, hit Queue.

- Rage to Vulnerability (Italian American, on the edge)
- Terror (sobbing, hyperventilating)
- Villain (the laugh comes first)
- Irish Woman, Dry Wit (sardonic Dublin)
- Eulogy (heavy with grief)
- Rain and Thunder (SFX plus urgent shouting)
- Italian Cooking Show (kitchen SFX plus enthusiasm)
- Old Male Storyteller (fireside, gravelly warmth)
- British Woman, East London Rage (screaming Cockney fury)
- Kid Explaining Dinosaurs (eight year old authority)
- Terrified Whisper (hiding, breath catching)
Pick “Custom” to write your own from scratch.
Writing your own voice
Three fields drive the performance.
Voice description. Age, timbre, accent, delivery style. This is where you establish who is speaking.
Male, mid 50s. Refined Central European accent with an Austrian tinge.Warm baritone that turns cold in an instant. Cultured, articulate,dangerously calm.Action tags. Opening cues, one per line. Sets the overall performance for the speech.
He smiles as he speaks, without warmthSpeech text. The words to say. Use inline [bracketed cues] for mid speech performance direction. Each bracket becomes a stage direction the model performs at that exact position.
You know, [he lets out a soft, dry laugh] I've always found politenessto be such a charming way of holding a knife. [His voice drops, suddenlyintimate] And all the while, their hands are already reaching for you.Available bracket cues include laughs, whispers, voice cracks, gasps, sobs, sighs, pauses, mood shifts, emphasis, and breath sounds. Anything you would write as a stage direction in a script.
Voice cloning from real audio
Use the built in LoadAudio node, wire it into Scenema Audio VAE Encode, and connect that node’s output to the Generate node’s ref_latent input.

Reference clips are capped at 20 seconds. Longer does not improve quality.
Twelve languages
English, Spanish, French, German, Italian, Portuguese, Japanese, Korean, Chinese, Hindi, Arabic, Swahili. Pick from the dropdown, then write the speech text in the target language.
Hardware
Minimum 8 GB VRAM. Tested end to end on RTX 3070 (8 GB) and RTX 4090 (24 GB). Generation runs up to 2x realtime.