ComfyUI Egregora Audio Super Resolution

A focused audio toolkit for ComfyUI: upscale, enhance, and evaluate audio quality with a clean, practical workflow. This pack is built for real-world use: minimal setup, clear node purposes, and tools to verify results.

---

Project scope (what this is and is not)

What it is:

• A set of audio enhancement nodes (FlashSR + Fat Llama) plus evaluation tools (ABX, loudness, null tests).
• Designed to help you improve low-quality audio and measure changes reliably.

What it is not:

• Not a magical "increase bitrate" tool. It enhances signal content and writes a new file at a chosen format/bitrate.
• Not a replacement for professional mastering. Think of it as an audio cleanup/boost stage.

---

Nodes overview (what each one does)

1) Audio Super Resolution (FlashSR)

Purpose: Diffusion-based upsampler aimed at musical content. It resamples internally to 48 kHz and can resample output back to your target rate.

Best for:

• Low to mid quality music or wideband content
• Improving detail and clarity in band-limited audio

Inputs:

• audio (AUDIO)
• lowpass_input (BOOL) gentle LPF before inference
• output_sr (48000 / 44100 / 96000)

Outputs:

• One AUDIO buffer

Use case:

• Feed an audio file node -> FlashSR -> Preview Audio

---

2) Spectral Enhance (Fat Llama GPU)

Purpose: Iterative spectral enhancement using CuPy on GPU.

Best for:

• Noisy or compressed audio
• Sharpening "sparkle" and spectral detail

Inputs:

• target_format (wav / flac)
• max_iterations (higher = more aggressive, slower)
• threshold_value (controls spectral gating)
• target_bitrate_kbps (target write bitrate)
• toggle_normalize (on by default)
• toggle_autoscale (on by default)

Outputs:

• One AUDIO buffer

Use case:

• Audio -> Fat Llama GPU -> Preview

---

3) Spectral Enhance (Fat Llama CPU/FFTW)

Purpose: CPU fallback using FFTW. Same idea as GPU but slower.

Use case:

• When you don´t have CUDA/CuPy

---

4) Enhance Extras

Purpose: Denoise, dereverb, and codec tools you can chain in front of FlashSR or Fat Llama.

Includes:

• RNNoise Denoise
• DeepFilterNet 2/3 Denoise
• WPE Dereverb
• DAC encode/decode

---

5) Eval Pack

Purpose: Measure loudness, distortion, and quality.

Includes:

• Loudness meter (LUFS approx)
• Gain match (LUFS/RMS)
• ABX preparation/judge
• Spectral metrics (SI-SDR, LSD)
• High quality resampler

---

6) Null Test Suite

Purpose: See exactly what changed between A and B by aligning and subtracting signals.

Includes:

• Alignment (GCC-PHAT)
• Gain match
• Null output and plots

---

How to combine nodes (common workflows)

Clean + enhance (recommended chain)

1. Denoise/Dereverb (Extras)
2. FlashSR (optional)
3. Fat Llama (light pass)
4. Eval Pack or Null Test to verify

FlashSR only

• Audio -> FlashSR -> Preview

Fat Llama only

• Audio -> Fat Llama -> Preview

---

Installation

1) Copy node pack

Place this folder into:

ComfyUI/custom_nodes/ComfyUI-Egregora-Audio-Super-Resolution

Restart ComfyUI once.

---

2) Install dependencies (recommended)

Use ComfyUI´s embedded Python:

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

Notes:

• Torch/torchaudio are not installed here to avoid breaking ComfyUI.
• On Windows, install.py installs NVIDIA CUDA runtime wheels for CuPy.

---

FlashSR repo and weights

The node auto-downloads the FlashSR inference repo on first use into deps/FlashSR_Inference/: https://github.com/jakeoneijk/FlashSR_Inference

However, FlashSR model weights are not included in this pack due to licensing/redistribution limits. The weights page does not state a license — download at your own discretion.

You must obtain the weights from the FlashSR authors or their official release and place them here: https://huggingface.co/datasets/jakeoneijk/FlashSR_weights

ComfyUI/models/audio/flashsr/
  student_ldm.pth
  sr_vocoder.pth
  vae.pth

Optional auto-download (if you host the weights in your own HF repo):

set EGREGORA_FLASHSR_HF_REPO=yourname/flashsr-weights

---

Troubleshooting (quick fixes)

FlashSR import issues

• The node auto-downloads deps/FlashSR_Inference/ on first use.
• If it fails, delete the folder and retry:

Remove-Item -Recurse -Force .\ComfyUI\custom_nodes\ComfyUI-Egregora-Audio-Super-Resolution\deps\FlashSR_Inference

CuPy / CUDA root not detected (Fat Llama GPU)

Run this in ComfyUI root:

python_embeded\python.exe -m pip install -U nvidia-cuda-runtime-cu12 nvidia-cuda-nvrtc-cu12 nvidia-cublas-cu12 nvidia-cufft-cu12 nvidia-curand-cu12 nvidia-cusolver-cu12 nvidia-cusparse-cu12 cupy-cuda12x

Numba needs NumPy 1.26 or less

python_embeded\python.exe -m pip install "numpy<=1.26.4"

---

License notes

• FlashSR inference code and weights are from upstream authors; check their repo for license status.
• Fat Llama packages are BSD-3-Clause (see PyPI).
• This integration is MIT (see LICENSE).

---

Changelog

• v0.2.1

  • FlashSR auto-bootstrap and clearer diagnostics.
  • Fat Llama CUDA path detection fixes for portable installs.
  • Fat Llama output scaling aligned with upstream behavior.
  • NumPy pinned to <=1.26.4 for Numba compatibility.

• v0.2.0 Added Enhance/Eval/Null toolsets; new installer + warmups.

• v0.1.0 Initial release: FlashSR SR node, Fat Llama GPU/CPU.