Skip to content

Refine (48 kHz) ​

Refine (48 kHz) brings a rendered song to 48 kHz and extends the missing top octave with a super-resolution model. It is one node with an optional model behind it, and it never touches the band the render already delivered.

No template uses Refine since 2026-10-21. All audio work after the render lives in the Plenio · Master block now, and its Audio HQ stage HQ 2 · Regenerate Highs regenerates the highs (with the same UniverSR model). Until then Refine ran in 3 · MiniMax · Song (preset 3 - MiniMax: pre 10 kHz, crossover 9.5 kHz, post 19 kHz) and sat bypassed in the other templates. The node stays in the library - Plenio · Refine (48 kHz) in the node library is the loader and the stage as one block - for a plain 48 kHz conversion (resample only) or your own experiments. The rest of this page describes the node.

The model always runs when it is connected - also for an input that already reaches the top of the band (decided 2026-09-28). PRE shapes only what the model sees, the crossover decides what the model may replace, and POST rolls its result off; nothing below the crossover changes. The report says so when the input was already full band.

The seed starts as fixed (its control after generate): the same seed gives the same result, and ComfyUI reuses the refined audio when nothing before it changed. Set it to randomize only to compare the model's variations.

What it does, step by step ​

  1. Measure the input's bandwidth (the highest frequency within 60 dB of the 1–4 kHz level).
  2. Condition the model's input (PRE): a steep low-pass - 0 takes the model's own training condition, -1 leaves the input unchanged, any other value is the low-pass in Hz - then resampling to the model's rate. This shapes only what the model sees, never the band you keep.
  3. Run the model chunk by chunk (5 s with 0.5 s overlap, crossfades, one derived seed per chunk).
  4. Join the original (converted to 48 kHz) and the model's output with a complementary crossover: LP(original) + HP(model), where the two filters sum to a pure delay. Below the crossover the result is the original, sample by sample; above it the model's content takes over.
  5. Optional POST roll-off (linear phase): 0 = off, else a low-pass in Hz (16, 19 or 21 kHz are the usual steps) if the added air is too bright.
  6. Contract: 48 kHz, exactly the input's duration, no normalisation. The report names the engine, the bandwidth before and after, the crossover, the loudness change, the peaks and the time.

A silent or very short input is only resampled, and the report says so.

The engines ​

EngineWhat it is
resample onlythe band-edge resampler (Kaiser, ≥ 100 dB stop band). No model file, no new content, bit-exact level.
modelthe connected Load Audio Model (super-resolution). The first engine is UniverSR (ICASSP 2026): vocoder-free flow matching on complex STFT coefficients, 4 ODE steps, 8/12/16/24 kHz → 48 kHz.

The model engine is mono inside: every channel is enhanced on its own - the stereo image is untouched.

No silent fallback: resample only is a visible engine choice. With model and no Load Audio Model connected, the node stops with a message that says what to connect.

Models and licence ​

FileFolderSizeLicence
pytorch_model.bin (UniverSR audio)models/audio_sr229 MBCC BY 4.0 (weights), MIT (inference code)

ComfyUI's missing-model dialog offers the download when you open 3 · MiniMax · Song; you can also put the file there yourself. The inference code is vendored in plenio/third_party/universr (pinned commit and per-file hashes in its NOTICE.md), so nothing is installed and nothing is downloaded by Plenio itself.

Presets and parameters ​

The preset list box sets PRE, POST and the crossover in one go; the crossover always sits 500 Hz below PRE (the rule since 2026-09-28):

PresetPRECrossoverPOST
1 - pre 6 kHz / post 16 kHz6 kHz5.5 kHz16 kHz
2 - pre 8 kHz / post 19 kHz8 kHz7.5 kHz19 kHz
3 - MiniMax: pre 10 kHz / post 19 kHz10 kHz9.5 kHz19 kHz
4 - pre 12 kHz / post 19 kHz12 kHz11.5 kHz19 kHz
5 - pre 14 kHz / post 21 kHz14 kHz13.5 kHz21 kHz
custom - the three fields belowpre_hzcrossover_hzpost_hz

A preset wins over the three fields; custom uses them as they are. The MiniMax template ships preset 3 and the matching numbers in the fields, so switching to custom keeps the same sound. The report names the preset that applied.

SettingDefaultMeaning
enginemodel in MiniMax, resample only elsewherewhich engine runs
crossover_hz0 = measured edge − 500 Hzwhere the model takes over from the original (used by custom)
sr_gain1.0level of the added band (lower it if the air is too strong)
presetcustoma prepared stage template (table above)
pre_hz0low-pass of the model's input: 0 = the model's training condition, -1 = none, else Hz (used by custom)
post_hz0linear-phase roll-off of the result: 0 = off, else Hz (used by custom)
seed0the model's seed (same seed, same result)

preset, pre_hz, post_hz and seed are optional inputs: an API-format prompt without them runs with the defaults above.

Provisional defaults: these values come from the design's re-evaluation of the legacy chain, not from measurements on your machine yet. The Refine block is marked experimental until the measurement study (L1) has decided them; the report marks the defaults as provisional.

The study (local) ​

tools/studies/sr_study.py measures the arms on your material, writes a JSON report and - with --pack <dir> - the blind A/B material (two level-matched files, key.json, README.md with the question). The commands and the decision rule are in the plan §4.5 and in CURRENT_STATUS.md's local checklist (L1). Nothing becomes a default before that verdict.

Limits ​

  • The model adds plausible high frequencies; it cannot recover what was never recorded. Above the crossover the result is the model's invention, not the artist's.
  • Refine has no loudness or peak management: peaks can rise with a boost. Plenio · Master follows it in every template.
  • It is the slowest stage in the chain; on a 16 GB card expect roughly real-time to a few times faster than real time with UniverSR (measured in L1).