Tag
The paper introduces VAE Rhythm, a periodic decoding trace left by 3D causal VAEs in AI-generated videos, and a training-free phase scan to detect it. It also presents CaDet, a detector trained on real videos paired with their VAE reconstructions, which outperforms baselines on DF26 and AIGVDBench across closed-source generators like Sora, Kling 3.0, and Seedance 2.0.
The paper presents PixVid, a method that lifts pretrained latent video diffusion models like Wan2.1-14B into pixel space by replacing the VAE decoder with a convolutional head, enabling higher-resolution, per-frame controllable video generation with competitive quality within ~1.3M training clips.
AMBIT是一个统一框架,通过将自回归场景规划与rectified-flow渲染耦合,实现第一阶Ambisonics(FOA)空间音频的生成与编辑,并利用FOA感知的连续VAE与on-policy自蒸馏来保持空间结构和声学保真度。
This paper investigates why the standard Gaussian prior N(0,I) on VAE latent spaces can fail under RL fine-tuning of latent diffusion models, showing that KL regularization compresses multimodal data into overlapping latent regions, and demonstrating a 71× improvement in collapse threshold using a Gaussian-mixture prior via a geometric mediator, with Lean 4-verified theory.
RoAR-VAE提出通过从受损隐变量中鲁棒地预测语义与重建信息来改进视觉tokenizer,在保持重建保真度的同时提升下游扩散模型的可学习性,平均生成基准提升2.19分,小文本重建精度较FLUX.1-VAE提升16.93个百分点。
LightVAE 提出一种视频自编码器压缩框架,通过对因果 Conv3D 层的精确重参数化,将空间响应与时间修正分离,并对时间修正施加低秩压缩,从而在保持生成质量的同时显著加速视频 VAE 解码(在 WanVAE 上最高达 2.24 倍吞吐提升)。
SILSA is a topology-aware 3D generation framework that replaces expensive voxel tokens with compact sliding-window slice latents and single-stage rectified-flow generation, using persistence-diagram/Betti-based slice-level topology supervision. It improves structural fidelity (PSNR +8.7%, Betti error -9.2%) while cutting tokens by 70%, memory by 40.4%, and inference time by 58.5%.
PAMI 提出一种 Part-Anchored Motion 框架,通过身体部位锚点投票机制和粗到细的生成层级(PamiGen 与 PamiRefiner)实现文本条件下的全身人-物交互生成,在 InterAct 数据集上较此前 SOTA 提升 14.5% 的接触召回率。
This study introduces a non-intrusive VAE-based latent-space adaptation method to correct CFD predictions of open tip clearance flow in compressor cascades, significantly reducing errors against sparse experimental observations.
PixSDS identifies VAE-induced pixel drift in latent score distillation sampling and proposes a gradient repair method that decodes latent SDS lookahead steps to guide pixel-space optimization, reducing artifacts in text-to-3D generation.
This paper studies a self-supervised task for generating single-cell gene expression vectors using an autoregressive transformer with a quantized VAE tokenizer. It reports scaling laws and a compute-optimal frontier for single-cell foundation models, with potential fine-tuning for perturbation prediction.
Kijai shares a quickly trained 2D tiny VAE for MiniMax-H3, intended to improve latent previews in ComfyUI. A better alternative trained by madebyollin is also linked.
Meshy T2 introduces a fast native mesh generation framework using flow matching and a vertex-set mesh VAE, achieving state-of-the-art geometric fidelity with end-to-end image-to-mesh generation in a median of 6 seconds, over an order of magnitude faster than autoregressive baselines.
OmniVAE is a jointly trained audio-video VAE that uses segment-level contrastive learning and feature distillation to align latent spaces, improving joint generation quality and synchronization in text-to-audio-video generation.
Proposes a VAE-based multi-task semantic communication framework for satellite-assisted autonomous driving, achieving significant bandwidth reduction while maintaining performance for traffic sign reconstruction and classification.
Introduces Cross-Space Distillation, a method to transfer knowledge from modern high-capacity diffusion models to compact student models with different latent spaces using a lightweight latent interface called Bridge, enabling quality improvements without modifying the student backbone.
This paper introduces TopVAE, a topology-optimized VAE that reduces 'dark areas' in molecular latent diffusion by making the decoder internalize structural and chemical constraints, achieving significant improvements in molecular generation quality.
Ideogram-4 model repackaged for ComfyUI, including fp8 scaled diffusion models, Qwen3VL text encoder, and FLUX VAE.
This paper introduces O-Voxel, a new sparse voxel representation for 3D generative modeling that efficiently handles complex topologies and appearance, and trains large-scale flow-matching models with 4B parameters to achieve state-of-the-art generation quality.