Research map / Image generation / Autoregressive and latent generation
Symbolic Music Generation with Transformer-GANs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add WGAN-GPen adversarial training with a pretrained BERT discriminator | Transformer-XL on MAESTRO MIDI V1 · MAESTRO MIDI V1 evaluation | CA: 0.838 -> 0.818 | Helped |
| Add WGAN-GPen adversarial training with a pretrained BERT discriminator | Transformer-XL on MAESTRO MIDI V1 · MAESTRO MIDI V1 human evaluation | preference: 57% prefer Transformer-GAN; 43% prefer Music Transformer | Helped |
| Add WGAN-GPen adversarial training | Transformer-XL on MAESTRO MIDI V1 · MAESTRO MIDI V1 validation | NLL: 1.74 -> 1.75 | Hurt |
| Use a pretrained BERT discriminator rather than a CNN discriminator (ablation) | WGAN-GPen Transformer-GAN on MAESTRO MIDI V1 · MAESTRO MIDI V1 evaluation | CA: 0.840 -> 0.818 | Helped |
| Add gradient penalty to RSGAN training (ablation) | RSGAN Transformer-GAN on MAESTRO MIDI V1 · MAESTRO MIDI V1 evaluation | CA: 0.862 -> 0.831 | Helped |
| Replace WGAN-GPen with RSGAN without gradient penalty (ablation) | Transformer-GAN on MAESTRO MIDI V1 · MAESTRO MIDI V1 evaluation | CA: 0.818 -> 0.862 | Hurt |
| Use randomly initialized BERT instead of pretrained BERT (ablation) | WGAN-GPen Transformer-GAN on MAESTRO MIDI V1 · MAESTRO MIDI V1 evaluation | CA: 0.818 -> 0.836 | Hurt |
| Freeze BERT embedding and layer 0 rather than freezing all BERT layers (ablation) | WGAN-GPen Transformer-GAN on MAESTRO MIDI V1 · MAESTRO MIDI V1 evaluation | CA: 0.818 -> 0.885 | Hurt |
| Freeze more layers of the pretrained BERT discriminator (ablation) | WGAN-GPen Transformer-GAN on MAESTRO MIDI V1 · MAESTRO MIDI V1 generated samples | CA: 0.843; 0.885; 0.859; 0.839; 0.818 | Mixed |
| Use TopK rather than Random sampling during generation (ablation) | Transformer-GAN on MAESTRO MIDI V1 · MAESTRO MIDI V1 generated samples | human evaluation | Hurt |
| Increase adversarial training sample length from 128 to 256, 512, or 1024 tokens (ablation) | Transformer-GAN on MAESTRO MIDI V1 · MAESTRO MIDI V1 generated samples | objective metrics | No clear effect |
Closest papers by meaning
- MuseGAN: Multi-track Sequential Generative Adversarial Networks for Symbolic Music Generation and Accompaniment (AAAI 2018 · similarity 0.92)
- Chunked Autoregressive GAN for Conditional Waveform Synthesis (ICLR 2022 · similarity 0.90)
- Presto! Distilling Steps and Layers for Accelerating Music Generation (ICLR 2025 · similarity 0.87)
- MuPT: A Generative Symbolic Music Pretrained Transformer (ICLR 2025 · similarity 0.87)
- SongGen: A Single Stage Auto-regressive Transformer for Text-to-Song Generation (ICML 2025 · similarity 0.87)
- N-gram Unsupervised Compoundation and Feature Injection for Better Symbolic Music Understanding (AAAI 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.