Research map / Image generation / Autoregressive and latent generation
Optimality of FSQ Tokens for Continuous Diffusion for Categorical Data with Application to Text-to-Speech
Vadim Popov, Wenju Gu, Tasnima Sadekova, Georgii Aparin, Assel Yermekova
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| replace CosyVoice2's 0.5B LLM text-to-token module with a 45M CDCD DiT on FSQ tokens | CosyVoice2 · SEED-TTS test-en | WER: 2.57% -> 2.00% | Improved |
| use true FSQ embeddings instead of perturbed embeddings (ablation) | CDCD-TTS · SEED-TTS test-en, 5 steps | WER: 3.10% -> 2.39% | Improved |
| use true FSQ embeddings instead of perturbed embeddings with many denoising steps (ablation) | CDCD-TTS · SEED-TTS test-en, 25 steps | WER: 2.03% -> 2.00% | No clear effect |
| keep FSQ embeddings but permute token ids (ablation) | CDCD-TTS · SEED-TTS test-en, 25 steps | WER: 2.00% -> 15.4% | Hurt |
| use RVQ 128-dim EnCodec tokens instead of FSQ 8-dim tokens (ablation) | CDCD-TTS · SEED-TTS test-en, 25 steps | WER: 2.00% -> 21.3% | Hurt |
| use true FSQ latent space instead of perturbed in toy unconditional generation (base2) (ablation) | CDCD · Toy sequences, base2 | log-KL: -8.68 -> -9.55 | Improved |
| use true FSQ latent space instead of perturbed in toy unconditional generation (base3) (ablation) | CDCD · Toy sequences, base3 | log-KL: -6.72 -> -7.16 | Improved |
| use FSQ codebook instead of random codebooks for optimal-model token accuracy (base3, 2D) (ablation) | CDCD latent space · Monte Carlo, 1000 random codebooks | prediction accuracy: 16.659% -> 17.052% | Improved |
| increase denoising steps from 5 to 25 for FSQ-original (ablation) | CDCD-TTS · SEED-TTS test-en | WER: 2.39% -> 2.00% | Improved |
| CDCD-TTS vs F5-TTS reference model on WER | CDCD-TTS · SEED-TTS test-en | WER: 1.83% (F5-TTS) vs 2.00% | Hurt |
Closest papers by meaning
- Continuous Diffusion Model for Language Modeling (NeurIPS 2025 · similarity 0.92)
- Efficient Perplexity Bound and Ratio Matching in Discrete Diffusion Language Models (ICLR 2025 · similarity 0.89)
- Masked Diffusion Models are Secretly Time-Agnostic Masked Models and Exploit Inaccurate Categorical Sampling (ICLR 2025 · similarity 0.88)
- Breaking AR’s Sampling Bottleneck: Provable Acceleration via Diffusion Language Models (NeurIPS 2025 · similarity 0.88)
- Discrete Diffusion Models: Novel Analysis and New Sampler Guarantees (NeurIPS 2025 · similarity 0.88)
- Flow Map Language Models: One-step Language Modeling via Continuous Denoising (NeurIPS 2026 · similarity 0.88)
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.
