🎧 Audio Super Resolution (FlashSR)

This is the node that gives the whole Egregora pack its name, and it's worth being clear-eyed about what it actually does. This is not a bitrate knob. You can't take a 128 kbps MP3 and "increase the bitrate" - that ship sailed at encode time, and every byte of missing high-frequency content is gone for good. What FlashSR does is infer that content back: a diffusion model looks at the band-limited audio you give it and hallucinates a plausible reconstruction of the frequencies that were cut off. If you've seen the TTS world rave about lifting 16–24 kHz model output to 48 kHz, that's this model family.

So when do you reach for it? Low-to-mid quality music, wideband voice, old recordings, or anything that sounds "dull" because the top end was chopped. It's genuinely good at that - FlashSR is well-established outside ComfyUI - but treat its output as restoration, not truth. It's also the one node in this pack where you should expect some fiddling before first run, because the weights aren't bundled with the pack.

How it works

The node resamples your audio to 48 kHz (FlashSR's design target), cuts it into 5.12-second chunks with a half-second overlap, runs each chunk through the model, and stitches it back with a Hann windowed overlap-add so you don't hear seams on longer files. A lowpass_input toggle runs a gentle low-pass filter before inference, which helps on very noisy input. Then, if you ask for a different output rate, it resamples the result back with an HQ cascade - so you get your original sample rate (or 44.1/96 kHz) out the other side.

The internals matter more than usual because of the setup cost: the node auto-downloads the FlashSR_Inference repo into deps/FlashSR_Inference/ on first use, but the three model weights - student_ldm.pth, sr_vocoder.pth, vae.pth - are not bundled, due to the upstream weights having no stated license. You grab them yourself and drop them in ComfyUI/models/audio/flashsr/ (you can also set EGREGORA_FLASHSR_HF_REPO=yourname/flashsr-weights if you host your own copy and want auto-download).

The inputs that matter

Only three required inputs, and two of them are trivial:

• audio - any AUDIO buffer from a load/preview node.
• lowpass_input - leave false for clean-ish sources; flip it on for noisy or aliased input where FlashSR might get confused.
• output_sr - 48000 / 44100 / 96000. Pick the rate your downstream wants. 48 kHz is the natural home turf.

Output is a single AUDIO buffer. Wire it into a save or preview node.

Install

ComfyUI Manager (search "Egregora Audio Super-Resolution") or:

cd ComfyUI/custom_nodes
git clone https://github.com/lucasgattas/ComfyUI-Egregora-Audio-Super-Resolution

Then install deps with ComfyUI's embedded Python and restart:

python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\ComfyUI-Egregora-Audio-Super-Resolution\requirements.txt
python_embeded\python.exe ComfyUI\custom_nodes\ComfyUI-Egregora-Audio-Super-Resolution\install.py

The pack deliberately avoids installing its own torch/torchaudio so it can't wreck your existing ComfyUI install - good call.

Where people get burned

The most common failure is the weights. If you run it and get "FlashSR weights missing," that's the licensing wall, not a bug - go fetch student_ldm.pth, sr_vocoder.pth, and vae.pth and place them in models/audio/flashsr/. If the auto-downloaded inference repo gets corrupted or import fails, delete deps/FlashSR_Inference and let it re-fetch. Expect the first run to be slow (model load + repo download); subsequent runs cache the model.

One honest caveat: as with any generative restoration, it can smooth over detail on already-decent audio. The Egregora pack includes a whole null-test suite precisely so you can verify whether FlashSR is actually helping or just sounding different - run the original and processed through the null test before you commit.