Mixing AI Stems
When the stems come from AI, the order of work has to change before the first fader moves.
Related
The problem
An AI stem pack looks clean at first. No room bleed, no mic leakage, no wrong notes. Solo a track and another set of problems appears. Parts leak into each other, the low end has been scooped by the separation model, every drum hit lands with the same weight, and the bridge is as full as the chorus. Running the recorded-stem order of EQ then compression makes these problems louder.
How it works
Bleed in a recorded stem comes from one room, so it arrives as delay and decay and reads as space. Bleed in an AI stem is spectral estimation residue; it sounds watery or hollow. A generative model returns a whole mix at once, and the parts are separated or generated in layers, so phase relationships are not guaranteed and low frequencies cancel easily. Dynamics are usually already limited, with weak sections as loud as strong ones. Those differences change the order of work, fix structure and level first, tone later.
Practice
Four passes. Listen to the whole and delete duplicate layers, since generators often produce two near-identical pads and keeping one removes a masking problem outright. Check polarity and phase with kick and bass in mono, flipping polarity or nudging timing if the low end thins. Repair dynamics with clip gain or volume automation before any compressor, using an expander ahead of compression and gating only the worst bleed. Only then do EQ and compression, high-passing non-bass parts at 80 to 120 Hz and cutting narrow bands at conflicts instead of boosting broad ones. Leave space until the end, because reverb hides bleed and blurs judgement. After each move, return to the full mix for ten seconds. Solo confirms a problem but should not decide it.
What to listen for
In the first movement of Beethoven's String Quartet No. 15, hear how four instruments of one family bleed into each other and how the midrange thins after separation. In the 86 BPM soul groove, hear whether every hit carries the same weight, since AI stems tend to flatten accents. In Ravel's Le gibet, listen to the low end between the bell and the chords, where separation residue collects. Judge in context first and use solo only to confirm.
Listen in the library
All tracks from midicn-lib — inline badge is the license tier (C1 commercial / C2 non-commercial / C3 study only).
Sources
- AI 分轨的串音与相位特征,依源分离与混音实践通识核对,未抄文本
- Stem-level gain staging and phase checks, based on general mixing practice, no text copied
Used for fact-checking only; all prose is original.