✨KLASS: KL-Guided Fast Inference in Masked Diffusion Models

📝 Summary:
KLASS accelerates masked diffusion model inference by using KL divergence to identify stable, high-confidence predictions. It unmasks multiple tokens per iteration, significantly speeding up generation and improving quality across text, image, and molecular tasks.

🔹 Publication Date: Published on Nov 7

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2511.05664
• PDF: https://arxiv.org/pdf/2511.05664
• Github: https://github.com/shkim0116/KLASS

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#DiffusionModels #GenerativeAI #MachineLearning #AIResearch #ModelAcceleration

✨Glance: Accelerating Diffusion Models with 1 Sample

📝 Summary:
Glance accelerates diffusion models with a phase-aware strategy using lightweight LoRA adapters. This method applies varying speedups across denoising stages, achieving up to 5x acceleration and strong generalization with minimal retraining on just 1 sample.

🔹 Publication Date: Published on Dec 2

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.02899
• PDF: https://arxiv.org/pdf/2512.02899

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#DiffusionModels #ModelAcceleration #LoRA #DeepLearning #GenerativeAI

✨StageVAR: Stage-Aware Acceleration for Visual Autoregressive Models

📝 Summary:
StageVAR accelerates visual autoregressive models by recognizing early stages are critical while later detail-refinement stages can be pruned or approximated. This plug-and-play framework achieves up to 3.4x speedup with minimal quality loss, outperforming existing methods.

🔹 Publication Date: Published on Dec 18

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.16483
• PDF: https://arxiv.org/pdf/2512.16483
• Github: https://github.com/sen-mao/StageVAR

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#ComputerVision #DeepLearning #ModelAcceleration #AI #NeuralNetworks

✨VisionTrim: Unified Vision Token Compression for Training-Free MLLM Acceleration

📝 Summary:
VisionTrim accelerates MLLMs by selecting dominant visual tokens and merging them with text guidance. This training-free framework improves efficiency without performance loss, addressing high computational costs from excessive visual data.

🔹 Publication Date: Published on Jan 30

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.22674
• PDF: https://arxiv.org/pdf/2601.22674

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#MLLM #VisionTokenCompression #ModelAcceleration #DeepLearning #TrainingFree

✨WaDi: Weight Direction-aware Distillation for One-step Image Synthesis

📝 Summary:
Diffusion model inference is slow. WaDi focuses on weight direction changes during distillation to accelerate models into efficient one-step generators. This achieves state-of-the-art quality with significantly fewer parameters and broad versatility.

🔹 Publication Date: Published on Mar 9

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.08258
• PDF: https://arxiv.org/pdf/2603.08258
• Github: https://github.com/gudaochangsheng/WaDi

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#DiffusionModels #ImageSynthesis #ModelAcceleration #DeepLearning #AIResearch