arXiv:2605. 13974v2 Announce Type: replace-cross Abstract: Diffusion Transformers (DiTs) and related flow-based architectures are now among the strongest text-to-image generators, yet the internal mechanisms through which prompts shape image semantics remain poorly understood.
WithinVAE is a compact variational visual autoencoder that learns semantic structure internally using a sequential bottleneck and self‑distillation from an exponential‑moving‑average copy of itself. It achieves a four‑fold compression ratio on ImageNet‑1K, improving reconstruction metrics such as PSNR by 1.36 dB over RPiAE while using half the latent capacity. The model also boosts linear probing performance, demonstrating that semantic gains arise from self‑distillation rather than external supervision.
arXiv:2608.25302v2 Announce Type: replace Abstract: Guided depth super-resolution (GDSR) typically extracts RGB guidance features through convolutional hierarchies, inheriting their fine-to-coarse bi...
FADRA is a frequency‑aware diffusion framework designed for video face restoration (VFR). It combines a pre‑trained text‑to‑video diffusion model with lightweight LoRA adapters and a Low‑Quality Pixel‑Alignment Feature Fusion module to adapt the generative prior to VFR. The method introduces a Repeated Residual Adaptation Head that refines residuals step‑wise using low‑quality latent cues, and employs a Frequency‑Aware Loss to supervise multiple spectral bands, improving facial detail recovery and temporal consistency.
arXiv:2610.35515v1 Announce Type: new Abstract: Diffusion-based real-world image super-resolution suffers from limited reconstruction fidelity due to information loss in VAE-based latent representati...
Video face restoration (VFR) aims to recover high-quality and temporally consistent facial details from severely degraded video sequences; however, existing methods still struggle to balance spatial fidelity and temporal coherence under complex degradations. To address this, we propose FADRA, a frequency-aware diffusion framework with iterative residual adaptation specifically tailored for robust VFR.
WAVE introduces a multi-level discrete wavelet transform (ML‑DWT) to reverse the typical fine‑to‑coarse bias in guided depth super‑resolution. By consuming wavelet sub‑bands and semantic tokens in reverse order, it separates structure and detail reconstruction, applies semantic gating to high‑frequency bands, and fuses modalities via an invertible coupling mechanism. Experiments on multiple benchmarks show that WAVE matches or outperforms existing methods, especially at high upsampling factors where low‑resolution depth has minimal structure.
AQUA-Net is a lightweight underwater image enhancement model that fuses spatial, frequency, and illumination information through a residual encoder‑decoder architecture with dual auxiliary branches. The frequency fusion encoder incorporates Fourier‑domain cues to preserve textures and details, while the illumination‑aware decoder applies adaptive exposure correction based on a learned illumination map. Experiments on benchmark datasets and a new high‑resolution Mediterranean Sea video set demonstrate that AQUA-Net achieves state‑of‑the‑art performance with fewer parameters, showing strong generalization and robustness for real‑world underwater imaging.
The paper introduces Composite Subspace Flow Matching (CSFM), a representation learning method that minimizes variance by inverting a deterministic degradation curriculum across spectral, color, and illumination subspaces. CSFM uses hard temporal style truncation to constrain an encoder, encouraging it to ignore transient style artifacts while preserving persistent content. Experiments on CIFAR-10, CIFAR-100, and STL-10 demonstrate that CSFM achieves competitive discriminative transferability and generative synthesis, outperforming standard optimal transport flows in linear probing and few‑shot classification while maintaining a smooth latent space.
arXiv:2603. 28762v2 Announce Type: replace-cross Abstract: Modern Text-to-Image (T2I) diffusion models have achieved remarkable semantic alignment, yet they often suffer from a significant lack of variety, converging on a narrow set of visual solutions for any given prompt.
The paper introduces Chameleon, a two‑stage training framework for cross‑domain image compositing that separates style and content representations. It first trains a ChameleonEncoder using Joint Hard Contrastive Learning to disentangle style and content, then applies Spatio‑Temporal Attention Gating within a diffusion transformer to stylize the foreground while preserving its identity. The authors also release ChameleonDataset, the first large‑scale training set for cross‑domain compositing, and demonstrate that Chameleon outperforms existing in‑domain, cross‑domain, and commercial models in both plausibility and stylistic fidelity.
The paper introduces Modality‑Specific Frequency Distillation (MSFD), a continual learning framework for video deepfake detection that separates spatial, temporal, and spatiotemporal features in the frequency domain. By preserving each modality independently and applying a cross‑modality decorrelation loss, MSFD adapts to new forgery patterns while maintaining performance across diverse continual deepfake video scenarios. Experiments demonstrate that this approach outperforms state‑of‑the‑art methods in both adaptation and retention.