Back to browse
Instrumental Model from Scratch (With Demo)

Instrumental Model from Scratch (With Demo)

by day6·Feb 18, 2026·1 point·0 comments

AI Analysis

●●SolidWizardryNiche Gem
The Take

The architecture is the project's real showpiece: a 72-band non‑uniform band-split BiMamba U‑Net that uses Mamba scans for O(T) memory and interleaved attention in the bottleneck to mix cross‑frequency context — a clever tradeoff between temporal efficiency and global attention. The author ships a runnable demo and an explanatory write-up so you can reproduce the approach, but it's clearly hobby-scale (≈1k songs trained, single home PC queue, slow cold starts), so expect experimental results rather than SOTA separation or instant throughput.

Category
Target Audience

Audio ML researchers, music tech hobbyists, curious engineers and anyone wanting quick instrumental stems of albums

Post Description

Wanted to try vibe coding a model one day when Opus 4.5 started becoming pretty good. Wrote up an intuitive explanation which should enable someone to generate similar code!

New conversions queue up and go through my one PC at home, so a disclaimer that new imports might be slow. I've also temporarily commited to the bit of having everything scale to $0, so I have lambda-ified my API server and deleted my EC2 instance so there could be a 0.1 second cold start. Mostly a hobby project with 0.4 users but please feel free to use it to listen to any album you'd like the instrumental version of!

Similar Projects

AI/MLMid

Custom ML model – if Spotify was instrumental

This ships a tidy, browsable library of on-demand vocal removals backed by a custom 'BiMamba UNet' trained on a curated ~500-song dataset — the interface actually feels built for music discovery rather than a bare demo. Don't expect SOTA separations (the author trained on one A40 for days and is honest about limits), but the vibe-coded dataset and integrated conversion pool make it a fun, useful tool for quickly grabbing instrumentals.

Niche GemShip It
day6
103mo ago
AI/MLMid

Trained YOLOX from scratch to avoid Ultralytics (iOS aircraft detect)

This is a practical, no-nonsense play: someone trained YOLOX from scratch, released MIT-licensed weights, and packaged a path toward running it on iOS. The value is procedural — dataset curation, training recipe, and an export/convert-for-iOS pipeline — but it's not a conceptual breakthrough; I'd like to see clear mAP numbers, model size and on-device latency benchmarks before recommending it for production.

Niche GemWizardry
auspiv
303mo ago
Education●●●Banger

How-to-Train-Your-GPT

Build a LLaMA-style model from scratch with zero ML prerequisites or math.

CozyBig Brain
RaiyanYahya
101mo ago