AI Paper Atlas

Research map / Image generation / Fast diffusion sampling

PriorGrad: Improving Conditional Denoising Diffusion Models with Data-Dependent Adaptive Prior

Sang-gil Lee, Heeseung Kim, Chaehun Shin, Xu Tan, Chang Liu, Qi Meng, Tao Qin, Wei Chen, Sungroh Yoon, Tie-Yan Liu

ICLR 2022 · Image generation › Fast diffusion sampling · approach: Denoising diffusion probabilistic models · 23 citations (OpenAlex)

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 changedWhereMeasuredEffect
Use a conditional data-dependent Gaussian priorDiffWave vocoder · LJSpeech testLS-MAE: 0.5264 -> 0.5048Improved
Use a conditional data-dependent Gaussian priorDiffWave vocoder · LJSpeech testMR-STFT: 1.0920 -> 0.9976Improved
Use a conditional data-dependent Gaussian priorDiffWave vocoder · LJSpeech testMOS: 4.12 ± 0.08 -> 4.25 ± 0.08Improved
Use a conditional data-dependent Gaussian priorDiffWave vocoder · LJSpeech test, 1M steps, 1.23M parametersMOS: 3.90 ± 0.09 -> 4.02 ± 0.08Improved
Use phoneme-level data-dependent Gaussian priorDiffusion acoustic model · LJSpeech acoustic model, 60K steps, 10M decoder parametersMOS: 3.84 ± 0.10 -> 4.04 ± 0.07Improved
Use phoneme-level data-dependent Gaussian priorDiffusion acoustic model · LJSpeech acoustic model, T_infer=2MOS: 2.80 ± 0.17 -> 4.25 ± 0.08Improved
Reduce decoder capacity from 10M to 3.5M parameters while retaining the PriorGrad prior (ablation)PriorGrad acoustic model · LJSpeech acoustic model, 300K stepsMOS: 4.09 ± 0.08 -> 4.08 ± 0.07No clear effect
Replace the fixed data-dependent prior with a jointly trainable Gaussian prior (ablation)PriorGrad acoustic model · LJSpeech acoustic model, 300K steps, 10M decoder parametersMOS: 4.09 ± 0.08 -> 3.32 ± 0.12Hurt
Use voiced/unvoiced labels instead of frame-level spectral energy to construct the vocoder prior (ablation)PriorGrad vocoder · LJSpeechMOSHurt
Use phoneme-level statistics instead of frame-level spectral energy to construct the vocoder prior (ablation)PriorGrad vocoder · LJSpeechMOSHurt
Use T_infer=6 instead of T_infer=50 (ablation)PriorGrad vocoder · LJSpeech testMOS: 4.33 ± 0.07 -> 4.20 ± 0.08Hurt
Use T_infer=6 instead of T_infer=50 (ablation)PriorGrad acoustic model · LJSpeech acoustic modelMOS: 4.39 ± 0.08 -> 4.29 ± 0.07Hurt

Later papers that compared with this method

4 later accepted papers report PriorGrad as a baseline in their results tables; 4 of them report a better number on most of the results they share with it.

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy