Apply for a GPU community grant: Academic project

#1
by Xuzheng - opened

Community GPU grant request โ€” SymphonyGen (ISMIR 2026 demo)

Hi HF team ๐Ÿ‘‹

I'm the first author of SymphonyGen: 3D Hierarchical Orchestral Generation with Controllable Harmony Skeleton, accepted at ISMIR 2026. This Space is the official interactive demo, and I'd like to request a ZeroGPU community grant so the research community can try the model without any setup.

What the demo does

SymphonyGen generates 32-bar orchestral pieces conditioned on a beat-quantized, multi-voice harmony skeleton โ€” a "short score" the user can sample from our harmony model, upload directly, or have analyzed from any MIDI file (re-orchestration). The demo exposes the paper's main knobs: model variants (GRPO-reinforced with a CLaMP 3 acoustic reward vs. pretrained), dissonance-averse sampling weights, and instrument constraints. Audio previews are rendered server-side with headless MuseScore, the same renderer used in the paper.

Fully open

Everything needed to reproduce the demo โ€” code, weights, and this Space โ€” is public.

Why a GPU grant, and why the fit is good

This is an academic side project with no hosting budget. The models are small (124M music model + 87M harmony model, ~2 GB total in fp32), and one generation takes on the order of a minute on a modern GPU versus tens of minutes on CPU โ€” a GPU makes the demo usable, but the footprint is tiny by ZeroGPU standards.

The Space is already built for ZeroGPU: models are preloaded on CUDA at module level, inference runs under @spaces.GPU(duration=120), audio rendering stays on CPU, and the public API is disabled with a single-concurrency queue, so GPU time is only consumed by real interactive use.

We expect traffic mainly from the ISMIR community, music-AI researchers, and composers curious about controllable symphonic generation. Thanks a lot for considering this โ€” happy to adjust anything about the Space's configuration if it helps.

Sign up or log in to comment