TASTE checkpoints

Model weights of "TASTE: Taxonomy-Aware Symbolic Piano Generation from Textual Description" (paper, code, dataset). TASTE generates solo-piano MIDI from six tags, a section plan and a free-text imagery description.

Path	Model	Size
ar/high_quality	TASTE (AR), final stage: the main model	2.4 GB
ar/pretrain, ar/large_public, ar/curated	TASTE (AR) after the earlier stages	1.7 / 2.4 / 2.4 GB
dit/high_quality	TASTE (DiT), final stage (EMA weights)	1.5 GB
codec/window_vae	window-VAE codec of the DiT, with its token vocabulary	351 MB
probes/structure_probe	section-structure probe	7 MB
probes/tag_probes	large-public tag probes and imagery retrieval bank	254 MB
transcription	Transkun audio-to-MIDI weights	54 MB

Usage

hf download Itsuki-music/TASTE --include "ar/high_quality/*" --local-dir $TASTE_CKPT_ROOT
python -m taste.inference.sample_ar --model-dir $TASTE_CKPT_ROOT/ar/high_quality --prompts-json prompts.json --output-dir out

The prompt format, the DiT sampler and the training stages are described in the code repository.

@misc{yao2026taste,
  title         = {TASTE: Taxonomy-Aware Symbolic Piano Generation from Textual Description},
  author        = {Mingyang Yao and Jingyue Huang and Ke Chen and Taylor Berg-Kirkpatrick and Shlomo Dubnov},
  year          = {2026},
  eprint        = {2610.15960},
  archivePrefix = {arXiv},
  primaryClass  = {cs.SD},
  url           = {https://arxiv.org/abs/2610.15960}
}