Step_Audio_EditX_SM_KSampler

Most ComfyUI audio packs only generate speech - prompt in, wav out. This node is the one that edits audio you already have: flip a recording from flat to angry, whisper it, make it sound older, add laughter or a sigh, clean up noise, even respeak the whole thing in someone's cloned voice. It's the ComfyUI wrapper around StepFun's Step-Audio-EditX, the LLM-based audio model that treats sound like a token sequence an LLM can rewrite - released open-source in November 2025 with its technical report on arXiv (2511.03601).

Honest take from people who've run it: the zero-shot voice cloning is genuinely good but a little bland on its own. The reason anyone actually reaches for this node is the editing second pass - the emotion, style, and paralinguistic controls are the draw, not the TTS. And expectations matter: several styles and emotions land so subtly you'll barely notice, while a few (laughter, whisper, anger) work really well. Don't expect a miracle every tag.

What the "sampler" actually does

The task dropdown is the whole personality of this node, and it has seven modes: clone, emotion, style, vad, denoise, paralinguistic, and speed.

• clone - zero-shot voice cloning. prompt_text is what's spoken in your input audio; target_text is the new sentence you want spoken in that voice. Give it 3–10 seconds of a voice and it'll say anything you type.
• emotion / style - edit mode. The model re-speaks your audio's content applying the tag you put in edit_info, e.g. happy, sad, angry (emotion) or whisper, child, older, radio, authority (style). The target_text field is ignored in edit modes - the words stay the same, only the delivery changes.
• paralinguistic - non-lexical sounds. This is the fun one: you type the tag into the text itself, on a second line, like 早上好,[Suprise-ah] ,吃了没 (or in English, insert [Laughter], [Breathing], [Sigh] at the point you want it).
• vad / denoise - cleanup modes. Voice activity detection and noise removal, no tag needed.
• speed - edit_info takes faster, slower, more faster, more slower.

n_edit_iter is how many passes the edit runs - 2 is the default and the README's sweet spot, 3 if you want the effect pushed harder. A big caveat the author is explicit about: the emotion tag has to roughly match the content of the audio. The model strengthens what's there; it can't invent crying-laughing from a flat line.

Inputs you'll actually touch

Three matter for everything, two are quality knobs:

• prompt_text - the transcription of what your input audio says (the loader's FunASR does this, but you can correct it; accuracy helps).
• edit_info - the emotion/style/speed tag. Default sad, which is a fun way to discover the node forgot to set it.
• offload - on by default; turn off only if you've got the VRAM and want speed.
• temperature - 0–1, default 0.7. Higher = more creative delivery, lower = more conservative.
• max_amplitude - the normalize-to-avoid-clipping ceiling (default 0.6). If input audio peaks above it, the node normalizes before processing.

Outputs are audio (wire it to PreviewAudio / SaveAudio, or into another audio node) and audio_path, a string pointing at the saved wav in your output folder - handy if you want the file path programmatically.

Install and running

Same pack as the model loader: clone smthemex/ComfyUI_Step_Audio_EditX_SM into custom_nodes, pip install -r requirements.txt, restart. Pin transformers==4.53.3 or you'll get silent output - that pin is load-bearing. See the Step_Audio_EditX_SM_Model article for the model layout: both Step-Audio-EditX and Step-Audio-Tokenizer must live under ComfyUI/models/SAEditX/.

The VRAM reality

This is a big model: StepFun quotes ~32GB of VRAM optimal at fp16, tested on an L40S. If you're under 16GB, keep offload on and expect it to be slow - it's the difference between this running at all and not. It's not a node you queue casually; it's a node you leave for a coffee break.