Audio Enhancement (DSRE)

LTX-2's big bet was synchronized audio-video generation, and its launch complaints were just as loud as its audio: thin, "tinny," sometimes metallic. Audio Enhancement (DSRE) is a single node that runs a whole mastering chain on that audio - de-essing, harmonic excitement, analog-style saturation, stereo width, dynamics - so you don't have to string together a dozen DSP steps yourself. People in the LTXV2 "metallic audio" threads literally recommend these nodes as the fix.

It's the other half of Urabewe's ComfyUI-AudioTools pack, built for his own LTX-2 GGUF workflows. The idea: place it between your audio decode and combine-video step, and it'll make the generated audio "at least a little better" - the author's words - without adding model downloads or a stack of extra nodes.

How it works

Under the hood it's a numpy/scipy DSP pipeline, roughly in this order: a de-esser for sibilance (4–10kHz), a low-pass above 16kHz to shave hiss, tanh saturation for "analog warmth," an 8-band harmonic exciter from sub-bass up to air, a psychoacoustic EQ boost focused on the 200–4000Hz range where voices live, gentle upward expansion for dynamics, then M/S stereo-width processing if the track is stereo. Finally it blends the processed audio back with the original by your enhancement_strength and peak-limits at 0.95. That blend cap matters: it's an aggressive chain, so strength is your sanity control.

The inputs that actually matter

• enhancement_mode - manual or auto_enhance. Auto analyzes frequency ratios (bass/mid/high/air) and an SNR estimate, then sets strength, bass, presence, warmth, and noise reduction itself. Manual knobs are ignored in auto. If you're new to this, just pick auto.
• enhancement_strength - 0 = bypass, 0.7 default, 1.0 = maximum. This is the master blend between processed and original.
• The tone stack - bass_boost (20–250Hz), presence_boost (2.5–8kHz clarity), warmth (tube-style saturation), dynamic_enhancement (1.0 = unchanged, higher = more expansion), stereo_width (only affects stereo audio; 1.0 = unchanged).
• enable_noise_reduction + noise_reduction_level - spectral gating for hiss, 1–10. At 10 it gets aggressive and can hurt quality.
• target_sample_rate - resamples the output; keep_original is default, 48000 fits most video.
• start_time / end_time / apply_to - full_track, selection_only, or v2v_mode.

v2v_mode is the reason this pack exists. It analyzes the 3 seconds before your selection as a reference, then tunes enhancement to match. The author's setup: set start_time where the original source ends and end_time at the end of the generated segment - the node compares and tries to make the LTX take-over blend in instead of announcing itself.

One honest caveat

harmonic_intensity doesn't actually do anything yet. It's threaded through the node's inputs and into the algorithm, but the shipped code never reads it - the exciter's harmonic content is fixed in the source. Crank it if you want; it's currently decorative. This pack is early (the author already shipped one "wrong version of the node" update), so check for updates if something misbehaves.

Installing it

From ComfyUI Manager, search for ComfyUI-AudioTools, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Urabewe/ComfyUI-AudioTools

Then restart ComfyUI. The requirements.txt is just resampy, soundfile, and librosa - the actual heavy lifting runs on scipy, which ComfyUI already ships, so there are no surprises and no model downloads. (The repo README is AI-drafted and even admits it; trust the code over its placeholder install URL.)

Common issues

• Sounds distorted - lower enhancement_strength to 0.5–0.6 and check the input isn't already clipping. This chain adds harmonics; feeding it a hot, clipped signal just makes a louder mess.
• No noticeable change - raise strength toward 0.8–0.9, or switch to auto_enhance which will boost where it detects weakness.
• V2V mode does nothing - the reference window needs real audio, so start_time should be at least 3 seconds into the file.
• Noise reduction makes things mushy - drop noise_reduction_level toward 1–3; balanced (5) is usually enough.