Adobe has introduced a new framework called AT-EDM, which stands for Attention-Driven Training-Free Efficiency Enhancement of Diffusion Models. This innovative approach aims to accelerate diffusion models by leveraging attention maps for runtime token pruning without the need for retraining. The AT-EDM framework has demonstrated a significant reduction in computational requirements, achieving approximately a 40% reduction in FLOPs while maintaining competitive generation quality across various aspect ratios.
[CV] Lumina-T2X: Transforming Text into Any Modality, Resolution, and Duration via Flow-based Large Diffusion Transformers https://t.co/rXcgKtmd4s - Proposes Lumina-T2X, a series of Flow-based Large Diffusion Transformers (Flag-DiT) that can transform noise to images,… https://t.co/5JL3LNCbNj
[CV] Distilling Diffusion Models into Conditional GANs M Kang, R Zhang, C Barnes, S Paris… [Pohang University of Science and Technology & Adobe Research] (2024) https://t.co/c5r5B9tFp0 - The paper proposes a method to distill a complex multi-step diffusion model into a… https://t.co/JDFZQGXO9J
Diffusion2GAN can generate a 512px/1024px image at an interactive speed of 0.09/0.16 seconds. By learning the direct mapping from Gaussian noises to their corresponding images, Diffusion2GAN enables one-step image synthesis. Paper: Diffusion2GAN: Distilling Diffusion Models into… https://t.co/VtG5yMawJZ
Lumina-T2X: Transforming Text into Any Modality, Resolution, and Duration via Flow-based Large Diffusion Transformers abs: https://t.co/Of841eZNPM code: https://t.co/fJc9wJfrMz Introduces the Lumina-T2X family of models, with the largest being a 7B DiT. Able to generate images,… https://t.co/aER8HdbRpu
Distilling Diffusion Models into Conditional GANs project page: https://t.co/M2z8vW1L56 abs: https://t.co/5n82ADW6qm Diffusion distillation treated as a paired image-to-image translation problem (noise → img). Train a simple conditional GAN with an LPIPS (perceptual) loss.… https://t.co/TcsVQub2fU
[CV] Imagine Flash: Accelerating Emu Diffusion Models with Backward Distillation https://t.co/NNB7EVUVRA - Proposes Imagine Flash, a novel distillation framework to accelerate diffusion models to generate high-quality images in 1-3 steps. - Introduces three key components:… https://t.co/37WCP4VIMb
[CV] Attention-Driven Training-Free Efficiency Enhancement of Diffusion Models https://t.co/WGhZpD2AQR - Proposes AT-EDM, a training-free framework to accelerate diffusion models by leveraging attention maps for run-time token pruning without retraining. - For single… https://t.co/rPY3rQdrob
📢Can you generate long multi-scene videos from existing text-to-video (T2V) generative models? Happy to introduce Time-Aligned Captions (TALC), a fundamental framework that allows multi-scene video generation without any further training 🤯 Paper: https://t.co/MfYIbVXppL 🧵👇 https://t.co/QDD65Q39hq
Attention-Driven Training-Free Efficiency Enhancement of Diffusion Models Wang et al.: https://t.co/mIouoN9q6a #ArtificialIntelligence #DeepLearning #MachineLearning https://t.co/GBtufFct6t
Diffusion models in text generation: a survey Diffusion models are a kind of math-based model that were first applied to image generation. Recently, they have drawn wide interest in natural language generation (NLG), a sub-field of natural language processing (NLP), due to their… https://t.co/pDGdfkPVWd
Diffusion models have shown remarkable performance in image generation in recent years. However, due to a quadratic increase in memory during generating ultra-high-resolution images (e.g. 4096 × 4096), the resolution of generated images is often limited to 1024×1024. This work… https://t.co/GbGVOGDcA7
Adobe presents AT-EDM: Attention-Driven Training-Free Efficiency Enhancement of Diffusion Models Achieves ~40% FLOPs reduction while enjoying competitive generation quality on various aspect ratios proj: https://t.co/8zd8KwYaJf abs: https://t.co/EvsR80FGD4 https://t.co/iQUEYBEJ4z