AI-generated tracks rarely have one problem. A Suno or Udio mix can arrive with shimmer, a ring on a vocal note and, once you split it into stems, drums behind the vocal. Run the right tools in the wrong order and each pass hides or undoes the last one’s work. This guide puts the stages in order, says which tool families cover each, and marks which parts of the order our own guides document and which are simply our default.
We make several of the plug-ins named below, so the sources are listed at the end and each comparison page states its limits.
The short answer
Work from structural problems to tonal ones. Start from the best source, fix gross errors, separate only if you need stems, remove bleed, clean generation or separation residue, treat resonances, then do your normal EQ, de-essing and compression, and compare with the original at matched loudness. Our guides document two parts of that order, bleed before residue and structural cleanup before EQ; the other stages are our default and we have not measured them.
The order, stage by stage
| # | Stage | What you do | Status of this step | Where to compare tools |
|---|---|---|---|---|
| 0 | Source | Start from the earliest lossless file; converting an MP3 to WAV does not restore what the MP3 discarded | Documented in WAV vs MP3 for AI music | — |
| 1 | Gross edits | Remove obvious clicks, count-ins and empty regions | Documented in our bleed guide | Any editor |
| 2 | Separation | Split into stems only if a problem is confined to one part | Our default | Separation models such as Demucs (archived), or a restoration suite’s separator |
| 3 | Bleed | Reduce the foreign instrument first and check the removed signal | Documented: bleed before residue | Stem bleed removal plugins compared |
| 4 | Residue | Clean the watery or metallic damage attached to the content you kept; on a vocal, start from the Vocal preset of de-artifact, our plug-in | Documented: residue after bleed | Eight artifact removers measured |
| 5 | Resonances | Treat rings that follow the notes | Our default | Dynamic resonance suppressors compared |
| 6 | Sustained vs hits | Optional: process sustained and transient material separately | Our default; placement not measured | Harmonic percussive separation tools |
| 7 | Normal processing | EQ, de-essing and compression | Documented: after the structural cleanup | Your usual chain |
| 8 | Check | Compare in the mix at matched loudness and keep a lossless master | Documented | — |
Why this order
Structural problems first, tonal problems after. Our bleed guide documents the order: remove identifiable instrument bleed first, then use artifact removal for the damage attached to the vocal. It also puts EQ, de-essing and compression after the structural cleanup; the reason is our reading, not a measurement: a tonal fix made before it chases a problem that is about to change. We place resonance treatment after residue because a suppressor reduces narrow peaks, while our guides describe generation residue as moving with the music; that placement is our reading of the rule, not a measurement.
What each stage can and cannot do
- Separation adds a problem of its own. Demucs, a separation model, was archived by its author on January 1, 2025 and its README says it is no longer maintained. iZotope lists Music Rebalance as the RX module that isolates vocals, bass and drums; its RX 12 edition table marks it for RX Standard and Advanced, not Elements.
- Bleed removal cannot always finish the job: dense accompaniment, reverb and backing vocals can make wanted and unwanted content inseparable.
- Residue removal can take wanted detail at strong settings. Our own testing found that at full strength on real, non-generated recordings it can take away genuine signal, so use it on AI-generated or AI-separated material and listen to what it removes.
- Bandwidth is a ceiling you cannot undo. The de-artifact manual describes positive Brightness as neural bandwidth extension: the band above 8 kHz crossfades to a regenerated one. That is an estimate, not the lost original; see the manual.
- Sound effects are a different problem; see why AI sound effects sound metallic or synthetic.
How to check each stage
- Solo what each tool removed. Bleed tools should remove the foreign instrument, residue tools the shimmer, suppressors the ring; if you hear words, melody or attacks, back off.
- Match loudness before every comparison.
- Compare against the previous stage, not only the original, so you know which pass helped.
- Check a dense passage and a sparse one before you commit to a chain.
- Keep a lossless copy of every stage’s output while you are still choosing.
Pick the stage you are stuck on
Each comparison states what the vendors say, where our tools are the wrong choice and how to test fairly.
Related guides
- Stem bleed removal plugins compared
- Dynamic resonance suppressor plugins compared
- Harmonic percussive separation tools compared
- Why do AI sound effects sound metallic or synthetic?
- What is an audio artifact?
Sources
- Demucs repository, github.com/facebookresearch/demucs, read October 20, 2026.
- iZotope RX 12 features and edition table, izotope.com/products/rx-elements, read October 20, 2026.
- Our own guides linked in the table, and the de-artifact manual.
FAQ
What should I run first when restoring AI-generated music?
Start from the best source you have, fix gross errors, separate only if you need stems, remove bleed, clean generation or separation residue, treat resonances, and do normal EQ, de-essing and compression last. Our guides document bleed before residue and structural cleanup before EQ; the rest of the order is our default and we have not measured it.
Do I need stem separation to restore an AI track?
No. Some tools work on a full mix. Separation lets you treat parts differently but adds bleed to deal with, so separate only when a problem is confined to one part.
Can restoration bring back frequencies lost to MP3?
No tool restores discarded information. A bandwidth-extension stage generates an estimated high band; in de-artifact, positive Brightness crossfades the band above 8 kHz to a regenerated one, which is an estimate rather than the original.
Will restoration make a track pass an AI detector?
That is not what restoration is for. Improving how a track sounds is a different task from changing how a detector classifies it, and we claim neither. Distributors set their own rules for AI-generated audio, so read them before you release.