arXiv cs.AI
·
18 hours ago
Hallo4D is a framework that uses large multimodal language models to detect and correct spatial and temporal inconsistencies in 3D and 4D content generation, such as duplicated structures and temporal jitter. The method employs a generation-detection-correction paradigm with LMM-based multi-model voting for consistency optimization without requiring retraining of existing models. The framework shows improvements over baseline methods across diverse 3D and 4D generation settings through techniques like motion-aware keyframe sampling and visibility pruning.
arXiv cs.AI
·
18 hours ago
Researchers developed DiT-Pruning, a post-training pruning method designed specifically for Diffusion Transformers that addresses limitations of existing pruning techniques by introducing customized saliency metrics and clustering-aware granularity. The method achieved only 0.001 CLIP score loss on FLUX.1-dev at 50% sparsity, compared to significant degradation from prior pruning approaches. This enables efficient deployment of diffusion models with reduced computational requirements while maintaining image generation quality.
arXiv cs.AI
·
18 hours ago
MedDiffuseMix is a diffusion-based image augmentation method that uses saliency maps to preserve diagnostically important regions while mixing lower-importance areas to improve medical image classification. The method was evaluated on four public datasets including chest radiographs and histopathology images, showing improvements in accuracy and F1-score over standard augmentation and baseline diffusion approaches. By constraining augmentation to clinically irrelevant regions, the framework maintains diagnostic evidence while increasing training data diversity for improved classifier performance.
arXiv cs.AI
·
18 hours ago
Researchers introduced PersGuard, a backdoor-based framework that protects against unauthorized personalization in text-to-image diffusion models by embedding protective backdoors into models before release. The framework uses three optimization objectives—backdoor behavior loss, prior preservation loss, and a novel backdoor retention loss—to ensure that fine-tuning on protected images triggers protective outputs while maintaining normal generation for unprotected images. PersGuard demonstrated superior privacy protection compared to perturbation-based defenses across gray-box, black-box, and multi-object protection scenarios.
