Scenema Audio Generate ⚡

Most local TTS will read your text out loud. This one performs it. Scenema Audio Generate is an expressive, diffusion-based TTS with zero-shot voice cloning, and its party trick is the <speak> prompt compiler: you type stage directions - [He laughs], [She whispers] - and the model actually performs them. On top of that you get twelve production presets, scene acoustics, language selection, automatic chunking, and an optional voice reference for cloning. It's the most "studio" thing in the local TTS stack, and it runs entirely on your own GPU.

The inputs that matter

The big three are model, clip, vae from Scenema Model Loader. Then the creative stack:

• preset - twelve presets that inject a voice, gender, scene and performance tags in one click: "Old Male Storyteller (fireside)", "Terrified Whisper", "Irish Woman, Dry Wit", "Villain (laughing menace)", even SFX ones like "Rain and Thunder". The speech_text is always used as entered, so presets are delivery, not content. Custom starts you from scratch.
• voice_description - your own voice spec, e.g. the default "Male, late 60s. Deep, gravelly. Slow and deliberate." Age, gender presentation, timbre, accent, delivery style.
• gender - grammatical gender for pronouns in the compiled prompt.
• speech_text - the text, with [bracketed cues] inline for mid-speech direction. Long text is auto-split at sentence boundaries.
• scene - acoustic environment (quiet room, reverberant hall, broadcast studio, café, rainy outdoors…), injected into the prompt. custom_scene overrides the dropdown with freeform text.
• action_tags - one stage direction per line, each becomes a direction the model performs.

Then: language (12 options), pace (default 1.5, a duration-budget multiplier - higher = slower; the validated default), and seed.

Cloning and the SeedVC pass

The optional ref_latent input (from Scenema VAE Encode, fed by Load Audio) gives zero-shot A2V cloning: the generation is conditioned on your reference voice's latent. It's the LTX A2V-style reference.

The second voice mechanism is the SeedVC post-pass: a fixed identity-consistency conversion applied to the output. identity_reference (an AUDIO) pins the identity for that pass, skip_vc turns it off entirely, and vc_steps / vc_cfg_rate (25 / 0.5) tune the conversion. The pass is seeded from the workflow seed, and here's a genuinely good behavior: Generate keeps the unpolished audio if conversion fails or returns anything other than a same-shape, same-length, audible waveform. You don't lose the clip on a post-process hiccup.

How it works under the hood

This is a port of ScenemaAI's ComfyUI-ScenemaAudio, rebuilt on ComfyUI's native LTX-AV machinery with no HuggingFace runtime downloads. The <speak> XML prompt compiler, the presets, sentence-boundary chunking with A2V voice chaining between chunks, and per-chunk trim/normalize before concatenation all carry over from the original. Output is 48kHz stereo.

Installing it

Part of the ComfyUI-GGUF-Loader pack under 🤖 CCTech/Scenema. ComfyUI Manager → search "ComfyUI-GGUF-Loader" → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/ChrisColeTech/ComfyUI-GGUF-Loader
cd ComfyUI-GGUF-Loader
pip install -r requirements.txt

The SeedVC pass needs the extras: place the seedvc, campplus, bigvgan, and whisper-small folders from the ChrisColeTech/scenema-audio extras repo under models/scenema-audio/extras. Skip that and either skip_vc stays on or the pass has nothing to run on.

Common issues

The classic failure is wanting the preset voice but typing speech text that fights it - presets set delivery, so a "Terrified Whisper" preset reading a grocery list still whispers a grocery list, just with conviction. For cloning, reference quality is everything: keep it clean, single-speaker, under 20 seconds (that's the VAE's hard cap). And the two voice inputs are not interchangeable: ref_latent is the A2V clone reference, identity_reference is the SeedVC identity - mix them up and the final voice won't match what you cloned. Long text: let the auto-chunking do its job; don't hand it one giant paragraph and expect a single breath.