AI Paper Atlas

Research map / Probabilistic and scientific modeling / Variational and generative sampling

Negative Binomial Variational Autoencoders for Overdispersed Latent Modeling

Yixuan Zhang, Jinhao Sheng, Wenxin Zhang, Quyu Kong, Feng Zhou

CVPR 2026 · Probabilistic and scientific modeling › Variational and generative sampling · approach: Regularizing VAE latent information

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
replace Poisson latents with negative binomial latents (MC-G variant)P-VAE · MNISTFID@5k: 105.3678 -> 79.6727Improved
replace Poisson latents with negative binomial latents (MC-G variant)P-VAE · CIFAR 16x16FID@5k: 60.3653 -> 40.2788Improved
replace Poisson latents with negative binomial latents (DS-C variant)P-VAE · CIFAR 16x16MSE: 0.0357 -> 0.0189Improved
replace Poisson latents with negative binomial latentsP-VAE · CelebA-64SSIM: 0.6354 -> 0.6329Hurt
use continuous-time reparameterization instead of Gumbel-Softmax with dispersion-sharing KL (ablation)NegBio-VAE · CIFAR 16x16SSIM: 0.6328 -> 0.8089Improved
use Gumbel-Softmax instead of continuous-time reparameterization with Monte Carlo KL (ablation)NegBio-VAE · CIFAR 16x16FID@5k: 67.2898 -> 40.2788Improved
use NegBio-VAE latents for MNIST fragmentation prediction vs P-VAENegBio-VAE DS-C · MNIST fragmentation prediction (N=200, latent 100)accuracy: 0.736 -> 0.811Improved
use NegBio-VAE latents for few-shot logistic regression vs best baselineNegBio-VAE · MNIST (20-shot)accuracy: 0.838 -> 0.865Improved
use NegBio-VAE latents for few-shot kNN vs C-VAENegBio-VAE · MNIST (20-shot kNN)accuracy: 0.791 -> 0.786Hurt
increase number of Monte Carlo samples from 5 to 25 (ablation)NegBio-VAE MC variants · MNISTMSE, SSIM, FID, KIDNo clear effect
increase beta scaling from small (0.2-0.4) to about 1.0 (ablation)NegBio-VAE · MNISTreconstruction vs FIDMixed
use stronger decoder (MLP or convolutional) instead of linear with linear encoder (ablation)NegBio-VAE · MNISTMSE, SSIM, FID, KIDImproved

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