A research team from New York University, MIT, and Google has recently proposed an innovative framework aimed at addressing the bottleneck of inference time scaling in diffusion models. This groundbreaking research transcends the traditional method of simply increasing the denoising steps, opening up new avenues for improving generative model performance. The framework primarily unfolds from two dimensions: one is leveraging feedback from validators, and the second is implementing algorithms to discover better noise candidates. The research team builds on a pre-trained SiT-XL model with a resolution of 256×256, maintaining 250 fixed denoising steps.

Microsoft Research has launched a powerful AI system called MatterGen that can generate new materials with specific properties, potentially accelerating the development of key technologies such as batteries and solar panels. The introduction of MatterGen signifies a fundamental shift in how scientists discover new materials. Unlike traditional methods that screen millions of existing compounds, MatterGen directly generates new materials based on desired characteristics, akin to how AI image generators create visuals from textual descriptions.

Recently, a line art coloring method called MangaNinja has garnered widespread attention, allowing users to color target line art based on reference images with just an input of the line art and the reference image. This technology is based on a diffusion model and focuses on reference-image-guided line art coloring, significantly enhancing the accuracy of coloring and interactive control capabilities. The research team employed two innovative designs to ensure precise transfer of character details. First, they introduced a patch rearrangement module to facilitate correspondence learning between the reference color images and the target line art. Secondly, they adopted a point-driven approach.

For a long time, efficiently generating high-quality, wide-angle 3D scenes from a single image has been a challenge faced by researchers. Traditional methods often rely on multi-view data or require time-consuming per-scene optimization, and they struggle with background quality and reconstruction of unseen areas. Existing technologies, when dealing with single-view 3D scene generation, often produce errors or distortions in occluded areas due to insufficient information, along with blurry backgrounds and difficulties in inferring the geometric structure of unseen regions. While regression-based models can synthesize new views in a feedforward manner, they struggle with complex...

In the field of artificial intelligence, ByteDance's commercial technology team presents the latest achievement, the Infinity model, which has become a new leader in the auto-regressive text-to-image generation domain with its outstanding performance and innovative technology. This newly open-sourced model not only surpasses Stable Diffusion 3 in image generation quality but also demonstrates significant advantages in inference speed. The core innovation of the Infinity model lies in its use of a Bitwise Token auto-regressive framework, which predicts the next level of resolution by +1 or -1.

Are you still troubled by unclear photos? A brand new image super-resolution tool called InvSR has emerged, allowing images to be instantly transformed into clear and sharp ones with minimal processing steps. The strength of this tool lies in its utilization of the rich prior knowledge embedded in large pre-trained diffusion models, achieving efficient and high-quality image resolution enhancement. The core technology of InvSR is its innovative partial noise prediction strategy. It cleverly constructs the intermediate state of the diffusion model as the starting point for the sampling process and employs a deep noise.

In the field of video analysis, the persistence of objects is an important cue for humans to understand that objects still exist even when completely occluded. However, current object segmentation methods mainly focus on visible (modal) objects, lacking the capability to handle non-modal (visible + invisible) objects. To address this issue, researchers proposed a two-stage method based on diffusion priors, Diffusion-Vas, aimed at improving the effectiveness of non-modal segmentation and content completion in videos, enabling the tracking of specified targets within the video and using diffusion models to complete the occluded parts.

Reinforcement learning has seen significant success in recent years, but its low sample efficiency limits its application in the real world. World models, as a type of environment generation model, offer hope for addressing this issue. They can serve as simulated environments to train reinforcement learning agents with higher sample efficiency. Currently, most world models simulate environment dynamics through discrete latent variable sequences. However, this method of compressing into a compact discrete representation may overlook visual details that are critical for reinforcement learning. Meanwhile, diffusion models have become prominent in the field of image generation.

Recent research from Alibaba's Tongyi Laboratory indicates that existing text-to-image Diffusion Transformer models already possess the ability to generate multiple images with specific relational contexts. With just a little 'guidance', they can achieve a synthesis of understanding and produce high-quality multi-image collections. Traditional Diffusion models are more like 'rote learning' students, requiring massive datasets to generate high-quality images. However, with the support of IC-LoRA, they become 'high-achieving students', capable of learning through analogy.

A recent study proposed a new method that leverages human attention to enhance the quality of AI-generated images. This method uses a saliency detector to identify the most important areas in the image and prioritizes the generation of these areas. Traditional image generation methods optimize the entire image uniformly, whereas the new method identifies and prioritizes more 'important' regions, similar to human attention. This approach can improve image quality and fidelity to textual prompts. Researchers utilized stable diffusion models and saliency detectors to generate images, and

At the 2024 Volcano Engine AI Innovation Roadshow, Volcano Engine President Tan Dai unveiled the latest Doubao · Video Generation Model. This model incorporates several advanced technologies, including the efficient DIT Fusion Computing Unit, which enables deep compression encoding of video and text; it also employs a novel diffusion model training approach to ensure consistency in generating multi-camera videos; in addition, the model integrates a deeply optimized Transformer architecture, significantly enhancing the generalization ability of video generation.

DiPIR is an innovative technology designed to address the technical challenges of seamlessly inserting virtual objects into real scenes. It achieves this by combining large-scale diffusion models with physics-based inverse rendering processes to accurately recover scene lighting information from a single image, allowing for the natural integration of virtual objects with their environments. The workflow of DiPIR includes constructing virtual 3D scenes, simulating interactions between objects and environments, and optimizing lighting conditions through diffusion models to ensure that generated images conform to real scenes. Its advantages include wide applicability, suitable for scenes under different lighting conditions, and the generation of highly realistic images.

Multimodal generative models are leading a new trend in artificial intelligence, focusing on the integration of visual and textual data to create multifunctional AI systems that perform various tasks, from image generation to understanding and reasoning across different data types. A key challenge is to enhance the capabilities of autoregressive (AR) models so they can generate high-detail images based on textual descriptions. While diffusion models excel in generating high-quality images, AR models lag behind in image quality, resolution flexibility, and multitasking capabilities. Researchers from Shanghai AI Lab and The Chinese University of Hong Kong have introduced Lum

A collaborative effort between the University of Alberta, the University of Toronto, and Huawei's Noah's Ark Lab has led to the development of a new method called TexGen, aimed at solving the challenge of generating realistic textures in the field of 3D modeling. This technology generates corresponding 3D textures from text descriptions, significantly outperforming existing techniques. The core involves a multi-view sampling and re-sampling framework combined with a pre-trained text-to-image diffusion model. This method effectively reduces discrepancies between views by maintaining a texture map with RGB parameters and updating it after each sampling, achieving view-consistent texture generation. An attention guidance mechanism has been introduced.

Stable-Hair is an innovative hair transfer technology designed to address challenges in handling diverse and complex hairstyles, providing users with an enhanced virtual hairstyle try-on experience. Utilizing diffusion models, it accurately transfers various real hairstyles onto user-provided facial images. The workflow consists of two phases: first, a 'baldness converter' removes hair from the image, followed by hairstyle transfer through three ....

In the realm of contemporary music creation, the application of AI technology is garnering increasing attention. However, the complexities of copyright issues and fair compensation for artists pose significant challenges for many startups in this field. Yet, Indian startup Beatoven.ai appears to have found a morally and responsibly sound approach to address these challenges.Mansoor Rahimat Khan, the co-founder and CEO of Beatoven.ai, is a professional sitar player hailing from a family with a se

Researchers from the University of Hong Kong and Tencent have proposed a new paradigm for multimodal recommendation systems called DiffMM, aiming to enhance the accuracy of short video recommendations. The system creates a graph containing user and video information and utilizes graph diffusion and contrastive learning techniques to better understand the relationship between users and videos, thereby achieving more accurate recommendations.The model method of DiffMM primarily includes three part

In recent research, a new method called Flash Diffusion has brought revolutionary breakthroughs to image generation technology. This method accelerates the generation process of pre-trained diffusion models by training predictive models to generate denoised multi-step prediction results in a single step.