I wanted to share my first completed production blending local open-weight model generation (YuE 2 via Vast.ai) with live tracked acoustic instruments.
Inspired heavily by Dva Brata’s Kafanski Blues and Morava, I wanted to capture that late-night smoke-filled kafana intimacy rather than modern polished styles.
Production Breakdown:
Foundation: YuE 2 (16-bar harmonic progression generated with acoustic guitar, accordion, and baritone vocal latents).
Live Overdubs: Tracked a Greek bouzouki and rhythm acoustic guitar through an AKG C214 into an Apollo Twin.
Lyricism: Written in Serbo-Croatian, focused on regret, autumn fog along the Sava river, and Turkish coffee going cold. Strict 10-syllable deseterac meter.
Stems: Split using Demucs v4 (htdemucs_ft) to carve space for the live bouzouki in the 1.2kHz–3.5kHz midrange.
Would love honest feedback from the room—especially on the vocal low-mid balance against the accordion!
Tarik, this is remarkably clean work. The dynamic range is so refreshing compared to the slammed -7 LUFS stuff floating around commercial streaming.
Mix notes:
320Hz build-up: Around the second verse (when the live bouzouki enters alongside the generated acoustic guitar stem), there’s a slight acoustic clutter between 280Hz and 340Hz. A dynamic EQ notch with a high Q (around -2.5dB on the generated guitar stem) would give your live bouzouki room to breathe.
Accordion stereo spread: The stereo image on the accordion is gorgeous. Did you use Mid/Side saturation or did YuE render that spatial depth natively?
The vulnerability in that vocal phrasing doesn’t sound like synthetic filler—you gave it human cadence and emotional weight. The deseterac meter works wonders for YuE’s prosody; the model doesn’t rush the consonants when the poetic foot has natural pauses.
Once you apply Clark’s EQ suggestion, polish the master to around -14 LUFS and submit it through the Open Stage intake form. We would love to host this in the player catalog.
Slušam ovo na slušalicama u Beogradu. Stvarno prenosi atmosferu kafane u novembru. The live bouzouki adds that acoustic imperfection that fools the brain into completely forgetting it started from model latents. Svaka čast!
Hvala svima! Clark, implemented that dynamic cut at 310Hz in FabFilter Pro-MB just now—the mud instantly vanished and the bouzouki pluck has proper snap now. And to answer your question: the accordion was widened using a subtle Haas delay on the sides only (+12ms), keeping the mid channel punchy. Henke, I will export the final master tonight!