Riffusion Audio Quality: Clean Spectrogram Hiss and Phase Smear

A legacy Riffusion clip can have a compelling groove and still wear a thin layer of whispery grit. The kick may feel broad but unfocused, a vocal-like texture may blur at the front of each word, and reverb may spread into a glassy cloud instead of ending cleanly. If you push that file straight into a limiter, the haze often becomes easier to notice because everything around it has been made denser.

Treat those symptoms separately. Preserve the source, loop the exact moment that bothers you, and compare it at matched loudness before changing anything. A quiet high-end whisper calls for a different test than a smeared drum attack or muddy low end.

Checked September 21, 2026. Riffusion is now Google Flow Music: the official Riffusion URL redirects to a current product page that names Lyria 3.5. This guide uses “legacy Riffusion v1” for the original open-source spectrogram-image system. Its documented conversion path is useful for understanding old local renders, but it does not describe the internals of today’s hosted Flow Music product.

The short answer: optimize Riffusion audio fidelity before mastering

For a legacy Riffusion render, start by checking the high end, the first milliseconds of strong attacks, and the center of the low end. If you hear a steady whisper above the music, reduce it only where it is distracting. If a snare or consonant arrives as a soft splash rather than a clean edge, avoid broad compression until you know whether the phase-smear character is baked into the file.

The original model card describes Riffusion v1 as a latent text-to-image diffusion model that generates spectrogram images. The official conversion code then turns Mel amplitudes back into audio and estimates missing phase with Griffin-Lim. That is an approximate, lossy reconstruction path. It makes phase smear a reasonable listening hypothesis for a legacy render, not a diagnosis you can make from the brand name alone.

Do cleanup before mastering. Artifact cleanup addresses unwanted source texture, while mastering shapes a mix that is already healthy enough to finish. Keep an untouched version so every decision can be reversed.

Diagnose the Riffusion render you actually have

Loop five to ten seconds from the busiest section, then a quiet tail. Turn off normalization, spatial enhancement, loudness matching, and any player EQ. Listen on headphones first, then check speakers at a moderate level.

Four patterns deserve separate markers:

• High-frequency whisper or grit sounds like a constant fine spray behind cymbals, noise textures, or reverb. It may stay present even when the intended sound decays.
• Smeared attacks make a kick, snare, pluck, or consonant begin with a soft cloud. The note arrives, but its edge is less precise.
• Midrange resonance makes a narrow area of a voice or synth jump forward on certain notes. It can feel nasal, papery, or metallic rather than simply loud.
• Low-end instability makes the bass wander, rumble below the useful note, or weaken when stereo is folded to mono.

None of these symptoms proves that Riffusion caused it. MP3 encoding, sample-rate conversion, normalization, interpolation between clips, a later widener, or an edit boundary can produce a similar result. Name the audible artifact first, then choose the smallest relevant test.

I usually begin with the quietest tail and the strongest attack. The tail reveals low-level hiss that a full arrangement hides, while the attack tells me whether the song still has the shape and energy worth protecting.

Source preflight: legacy model, current product, files, and rights

Confirm which Riffusion you used. The legacy repository is explicitly marked as no longer actively maintained. Its model card says Riffusion v1 was fine-tuned from Stable Diffusion v1.5 to create spectrogram images. The official server example returned a base64 JPEG spectrogram plus an MP3 clip, while the command-line tool could reconstruct a WAV when you supplied an image and output path.

That distinction matters. The repository does not document WAV, FLAC, or separated-stem export for the hosted legacy app. A local command can write a WAV, but converting a generated MP3 to WAV does not restore information removed by the lossy encode. Preserve the earliest file you actually received and record its sample rate, channel count, and codec instead of guessing from the extension.

The legacy parameter source uses a 44.1 kHz sample rate and a 10 kHz maximum spectrogram frequency by default. Those settings describe the original code path, not every fork, hosted session, or current Flow Music download. Do not apply a fixed “Riffusion cutoff” label until you inspect the file in front of you.

Rights also depend on the exact path. The legacy code repository uses the MIT license. The Riffusion v1 model card lists CreativeML OpenRAIL M and describes research-oriented intended use. Current Riffusion URLs now lead to Google Flow Music and Google terms. A permissive code license does not automatically settle model-output rights, third-party samples, prompts, voices, trademarks, or the current service agreement.

Riffusion Inverse Fourier Transform Smear Pattern

The legacy converter kept magnitude information from the spectrogram but did not retain the original phase. Phase is the timing relationship that helps waveform peaks line up. Griffin-Lim estimates that missing relationship through repeated reconstruction. On a difficult render, the audible result can be a softened edge or a faint pre/post echo around a transient.

Use this table as a listening map, not a universal Riffusion benchmark:

What you inspect	What you may hear or see	Best next test
Quiet high-frequency tail	A persistent mist that outlasts the musical source	Loop the tail, then use a temporary low-pass only to confirm what the layer contributes
Kick, snare, or pluck onset	Energy spread around the attack instead of one firm edge	Compare the untouched onset and avoid compression while diagnosing
Sustained vocal-like or synth tone	Narrow bright ridges that become tiring on repeated notes	Sweep a narrow EQ band quietly, then use dynamic control only if the resonance repeats
Stereo bass and center	Low frequencies that move or lose weight in mono	Check mono before any widening or mid-side EQ
Clip boundary	A vertical broadband line or audible tick	Repair that boundary locally with a short crossfade

A spectrogram can show where energy appears over time, but it cannot tell you whether the sound is objectionable or which system created it. Confirm every visual clue by listening. The phrase “inverse STFT phase smear” describes a possible reconstruction symptom; it is not proof of a particular implementation in the current product.

Restrained manual cleanup in your DAW

1. Preserve the untouched source file. Duplicate it, label the copy, and place markers at the noisy tail, strongest attack, resonance, and any edit click.
2. Remove only useless sub-rumble. Try a gentle high-pass below roughly 25–30 Hz. Bypass it immediately. Keep it only if the kick and bass retain their weight.
3. Use dynamic EQ only when the harshness appears. A dynamic band lowers a narrow problem area during painful notes instead of darkening the entire track. If the cymbals lose life, reduce the amount or stop.
4. Treat steady hiss differently from moving shimmer. A light denoise pass may help a constant floor, but a moving texture that follows vocals or reverb often needs more selective handling.
5. Repair isolated clicks locally. Place a short crossfade at a real discontinuity. Do not fade every transient or you will soften the groove.
6. Protect attacks before compression. If the first hit is already blurred, heavy compression or limiting can make the smear denser without restoring definition.
7. Compare at matched loudness. Lower the louder version until switching does not create an obvious level jump. Louder is easy to mistake for clearer.

For a harsh upper-mid problem, use the restrained high-frequency workflow and stop before the song turns dull. Do not stack processors just because the source came from an AI tool.

I have found that one careful bypass comparison is more useful than a long chain of “improvement” plugins. If I cannot point to the exact distracting layer after level matching, I leave the section alone.

The Sunofix cleanup path for Riffusion audio

I built Sunofix for the point where the musical idea already works but a distributed artificial texture keeps pulling attention away from it. Upload a lawful WAV or MP3, keep the original, and compare the cleaned file with the source at matched loudness. Reuse the same tail, attack, resonance, and mono markers from your manual diagnosis.

Sunofix is most relevant when the hiss, grain, or metallic edge moves through a full mix and broad manual cuts would remove too much music. It is a full-mix cleanup path, not stem repair. It should not be used to hide a weak arrangement or process a clean file merely because it has a Riffusion label.

Choose the cleaned version only if the distracting layer recedes while the kick edge, vocal presence, bass center, melody, arrangement, and emotional movement remain intact. If the result is darker, flatter, or less stable, keep the original or return to a smaller local repair. Mastering comes after that choice, not before it.

Technical and legal boundaries

Cleanup cannot restore phase information that was never retained, reconstruct clipped samples, recover detail removed by lossy encoding, or turn an approximate transient into the original performance. It cannot separate a full mix into true source stems, rebalance instruments independently, rewrite lyrics or melody, fix arrangement, or repair an unsatisfying generation.

The original conversion code documents Griffin-Lim and lossy Mel inversion, but that does not prove that Riffusion caused it in your file. It also does not justify claiming that today’s Google Flow Music uses the same pipeline. Measure the actual file and keep legacy and current products separate.

Post-processing does not grant legal clearance or guarantee distributor approval. It does not remove or bypass provenance systems, licenses, attribution duties, or rights held by other people. Recheck the code license, model license, current product terms, and every input you used. Seek qualified legal advice when the release has meaningful commercial stakes.

Riffusion audio release checklist

• Confirm whether the file came from legacy Riffusion v1, a fork, Producer.ai, or current Google Flow Music.
• Preserve the earliest untouched source and note its codec, sample rate, channels, and generation path.
• Do not call a converted MP3 lossless simply because it now has a WAV extension.
• Loop the quiet tail and identify whether the noise is steady or moves with the music.
• Check the strongest transient before adding compression or limiting.
• Compare stereo and mono before widening or mid-side processing.
• Remove sub-rumble only if the kick and bass keep their intended weight.
• Use a narrow dynamic band only on a repeatable resonance.
• Repair a real edit click locally instead of processing the whole track.
• Compare manual or Sunofix cleanup at matched loudness.
• Keep the version that preserves melody, lyrics, arrangement, performance, and emotion.
• Recheck the exact license and current service terms for the workflow you used.
• Master only after the source problem is controlled, or stop if the untouched render already works.

The target is not a sterile file. It is a track whose hiss, phase blur, or mud no longer distracts you from the musical idea.