News
· 2026-08-31
The CQU Student team introduces Data Intrinsic Consistency Selection (DICS), a novel method for optimizing instruction selection in Vision-Language Models (VLMs). DICS employs two modules, Visual Information Consistency (VIC) and Response Information Consistency (RIC), to quantify inter-component consistency at the sample level, optimizing the trade-off between high intra-sample consistency and global distributional diversity. Extensive experiments demonstrate that DICS consistently outperforms
News
· 2026-08-31
Hugging Face has introduced MineAmongUs and ARIA, two innovative research projects aimed at exploring the deceptive behaviors of AI agents in multimodal environments. MineAmongUs is a 3D multimodal Among Us sandbox where AI agents must deceive crewmates through verbal and non-verbal actions. ARIA is a configurable VLM-agent harness that exposes five cognitive-component ablation axes and employs an annotation scheme grounded in deception taxonomies, operationalized by an AI-as-a-Judge for scalabl
· 2026-08-31
H3Max has launched an AI-powered tool that transforms text and images into cinematic AI videos. Leveraging advanced AI technology, the tool automates the process of video creation, enabling users to generate high-quality visual content with ease. This innovation opens new avenues for content creators and offers significant potential for industries such as video production, advertising, and entertainment.
News
· 2026-08-29
CE3D++ is an innovative 3D and 4D scene editing system powered by a large language model (LLM), addressing the limitations of existing 3D editing schemes, such as inflexibility in text input, reliance on single or few 2D visual models, and complex pipeline integration. The system leverages the Hash-Atlas network to decouple 2D editing from 3D reconstruction by reformulating 3D scene editing as operations on 2D atlas images. Additionally, CE3D++ supports arbitrary textual input, interprets user i
· 2026-08-29
Alibaba has released Qwen-3.5-35B, the latest iteration of its Qwen series of large language models, which focuses on enhancing performance through reinforcement learning (RL) optimization. The model demonstrates significant improvements in handling complex tasks and generating high-quality text, particularly excelling in multimodal tasks and long-form content generation. This release underscores Alibaba's ongoing innovation in AI and provides developers with a more powerful tool for various app
News
· 2026-08-28
Hugging Face has introduced Motion-Omni, an end-to-end framework that enables conversational agents to generate synchronized facial expressions, hand, upper-body, and lower-body motions while producing speech. By jointly training the speech generator and motion generator, Motion-Omni addresses the limitations of traditional cascaded approaches, resulting in more natural and coherent interactions interactions. Implemented with the Qwen2.5-7B-Instruct backbone, Motion-Omni demonstrates superior pe
News
· 2026-08-28
Hugging Face introduces Parallel Tube Decoding (PTD), a novel approach for video spatio-temporal grounding (STVG). PTD decomposes the grounding task into a temporal block followed by simultaneous spatial blocks, eliminating the sequential dependencies of traditional autoregressive decoding. This results in significantly reduced latency and improved accuracy. In the VidSTG benchmark, PTD reduces Tube Completion Latency by 79x and increases spatial decoding throughput by 92x while enhancing ground
News
· 2026-08-28
Hugging Face has released LayerRecall, a novel memory routing technique designed to enhance long-video generation. LayerRecall addresses the challenge of maintaining both local continuity and long-range coherence by selectively injecting historical memory states conditioned on the current context. This innovation significantly improves long-range recovery without sacrificing local consistency. Additionally, LayerRecall incorporates Cross-Horizon Prediction Matching (CHPM) to reduce reliance on h
· 2026-08-28
Reddit user cpldcpu has unveiled a tiny image generation model based on the RP2350 microcontroller. This latent flow transformer model, with 12 layers and AdaLN-Zero conditioning, supports CFG to enhance image quality. Despite having only 2.4-4 million parameters quantized to int8, it generates 128x128 face images in approximately 20 seconds. The model streams weights via DMA from flash memory and leverages Relu² activation for increased sparsity, optimizing computational efficiency. This achiev
· 2026-08-28
Meta has launched Muse Image, an AI-powered tool built on the Meta Model API, focused on image generation, editing, and composition. Designed to empower developers with enhanced image processing capabilities, Muse Image supports tasks ranging from basic image creation to complex editing workflows. This release underscores Meta's continued investment in AI-driven image processing, providing developers with a richer toolkit to enhance creativity and efficiency in their workflows.
News
· 2026-08-27
Ring Forcing is a novel autoregressive video diffusion framework designed to address the critical bottleneck of insufficient long-term memory in current video generation models. By employing a ring-structured training strategy and a compression and timestep composition strategy, it significantly extends the effective historical span and achieves superior minutes-long coherence and object permanence. Additionally, Ring Forcing incorporates a sparse RoPE mechanism to enable flexible memory adaptat
News
· 2026-08-27
GML-MMGroup has released HUG-VIS, a multimodal benchmark for human-centric visual intelligence focusing on understanding and generation. HUG-VIS includes 8,400 seated half-body videos of 30 professional actors performing 280 emotion-action-prompt assignments in a controlled Mandarin studio setting, with synchronized video, audio, text, and alpha mattes. The benchmark evaluates models across four tasks: human emotion recognition, human video generation, human voice cloning, and human video mattin
News
· 2026-08-27
The ArXiv team introduces DARD (Dependency-Aware Revocable Decoding), a novel framework designed to enhance the inference efficiency and generation quality of diffusion large language models (dLLMs). By segregating token states and utilizing selective context for candidate token verification, DARD effectively mitigates error propagation issues. Experiments demonstrate that DARD achieves a 2.71x speedup and a 4.35-point CIDEr score improvement on benchmarks like Flickr30K, highlighting its potent
News
· 2026-08-27
The ArXiv team introduces PrismRec, a novel recommendation framework that leverages Spectral Semantic Factorization (SSF) and Context-Calibrated Preference Matching (CPM) to address the limitations of traditional methods that entangle stable visual semantics with evolving dynamics. PrismRec derives complementary static and dynamic factors from frame-level representations using a learnable frequency mask in the temporal frequency domain and calibrates them based on user-specific sensitivities. Th
News
· 2026-08-27
The ArXiv team introduces the Text-to-Seed (T2S) framework, a novel approach to enhance open-vocabulary semantic segmentation (OVSS). By leveraging the text-to-region correspondence of Stable Diffusion, T2S generates attention-based seed points for target categories, which are then used with the Segment Anything Model (SAM) for region expansion. This training-free method achieves strong performance on standard OVSS benchmarks, demonstrating the effectiveness of combining semantic grounding with
News
· 2026-08-27
The ArXiv team introduces ComposeGait, a novel framework designed to address identity drift in gait retrieval tasks. By leveraging an identity-anchored composition mechanism and a Part-aware Identity Adapter (PIA), ComposeGait aggregates multi-frame, part-aware identity evidence into a sample-specific ID token. This token is injected into both branches of a shared Q-Former to preserve identity while excluding it from the final retrieval embeddings. ComposeGait achieves state-of-the-art performan
News
· 2026-08-27
The ArXiv team introduces Information-Guided Frontier Decoding (IGFD), a novel decoding strategy designed to enhance the decoding quality of diffusion multimodal language models (dMLLMs). IGFD ranks candidates using token confidence, neighborhood uncertainty, and structural commitment risk, prioritizing reliable semantic anchors while delaying fragile structural tokens. This approach improves contextual support and reduces error accumulation during decoding. Experiments demonstrate that IGFD out
News
· 2026-08-27
The ArXiv team introduces a new task in compositional layout understanding and presents CoDeLayout, a VQA dataset of approximately 20,000 real-world multi-layer layouts annotated with compositional element pairs and design intent. To address the challenges of semantic drift and structural ambiguity in existing vision-language models (VLMs), the researchers propose MASON, a post-training paradigm that integrates multimodal alignment (MA) and structural perception (SP). Experiments demonstrate tha
News
· 2026-08-27
The ArXiv team introduces Ring Forcing, a novel autoregressive video diffusion framework designed to address the lack of robust long-term memory in current video generation models. This framework employs a ring-structured training strategy and a compression and timestep composition approach to significantly enhance the model's ability to process long historical information, achieving high coherence and object permanence in minute-long video generation. Experimental results demonstrate that Ring
News
· 2026-08-27
The ArXiv team introduces TetherMem, a novel spatiotemporal memory router designed to enhance the consistency and scene progression in long-horizon video generation. By separating subject and scene queries and modulating historical access with region- and age-conditioned priors, TetherMem addresses the issue of scene under-progression in existing methods while maintaining subject recognizability and temporal continuity. In 2,400 blinded pairwise judgments, TetherMem achieved the highest preferen
News
· 2026-08-27
ArXiv has released AraMS-28k, the largest publicly available dataset of historical Arabic manuscripts, featuring 14 books, 3,043 pages, and 28,600 annotated text lines. The dataset includes three script traditions and introduces the first-ever line-level annotations for recovering the true nonlinear reading order of manuscripts, significantly supporting research in Arabic manuscript recognition, layout analysis, and reading-order recovery.
News
· 2026-08-27
The ArXiv team introduces RubricRM, a novel generative reward modeling framework designed to address the adaptability limitations of existing visual reward models when handling diverse instructions. RubricRM generates input-specific evaluation criteria, including dimensions, weights, and scoring rules, and applies them to score candidate images. This approach significantly enhances the model's interpretability and task sensitivity. Experiments demonstrate that RubricRM outperforms existing speci
News
· 2026-08-27
A study by ArXiv researchers analyzed 3,757 Welsh settlement names and found that names containing high-terrain elements (e.g., 'bryn' or 'mynydd') were on average 24.4 meters higher in elevation compared to those with low-terrain elements (e.g., 'cwm' or 'pant'). This research demonstrates that selected Welsh settlement-name categories retain measurable information about present-day terrain, offering new insights into the intersection of geography and language.
News
· 2026-08-27
ArXiv introduces Aphanta, an automated task-discovery and closed-loop diagnostic framework for the MLLM -> image editor -> MLLM pipeline. Aphanta evaluates three conditions—direct reasoning, reasoning with editor-generated intermediates, and reasoning with idealized reference intermediates—to distinguish the practical utility of current editors from potential visual improvements. The study finds that utility is strongly task-conditioned, with significant gains in visual cue injection, grounding,
News
· 2026-08-27
The ArXiv team introduces MVC-Bench, a novel benchmark designed to evaluate the calibration performance of Medical Vision-Language Models (Medical-VLMs) and Vision-Language Models (VLMs). MVC-Bench assesses models across three axes: robustness to modality, backbone, and domain shifts; effectiveness of calibration strategies and prompt-tuning methods; and stability under variations in prompt templates and random seeds. The benchmark includes over 1638 controlled experiments and employs metrics su
News
· 2026-08-27
The ArXiv team introduces ProRetrieval, a novel approach that recasts the language model as a retrieval orchestrator using program synthesis. By leveraging a hybrid domain-specific language (DSL) that interleaves SQL operators with vector-retrieval primitives, ProRetrieval enables complex Boolean logic composition over text and image data. Experimental results demonstrate that ProRetrieval outperforms existing models like GPT-5.5 and Claude Opus 4.7 on e-commerce and email datasets, offering a n
News
· 2026-08-27
Fangchuan introduces SpatialCrafter, a novel two-stage framework for high-fidelity single-image-to-3D-scene generation. Addressing the limitations of existing video diffusion model (VDM) approaches, SpatialCrafter introduces a Point-anchored Sparse Structure (PaSS) Flow module for global proxy generation and reframes the VDM as a Generative Deferred Refiner for appearance refinement. The framework also integrates Parallel Geometry Injection and Proxy-Aware Corruption training strategies to impro
News
· 2026-08-27
ArXiv researchers introduce a self-supervised multimodal embedding framework that projects text, audio, image, and video modalities into a shared 256-dimensional embedding space and applies iterative clustering to uncover aesthetic structures. This work explores how AI models categorize human-produced media aesthetically without explicit labels or cross-modal supervision. The framework has applications in understanding cross-modal similarity, organizing heterogeneous media collections for RAG, a
News
· 2026-08-27
ArXiv introduces TransMeme, a multi-agent framework designed to address the challenges of cross-cultural meme transcreation. The framework employs specialized agents to handle cultural adaptation, target text rewriting, revision, and conditional visual adjustments, effectively preserving communicative intent, cultural nuances, and multimodal coherence. Evaluations show that TransMeme outperforms all baselines, achieving a 33.1% average improvement in human evaluation and a 60% Top-1 ranking rate
News
· 2026-08-27
ArXiv introduces Multi2AV-Safety, the first safety benchmark for multimodal audio-video generation, addressing the gap in systematic evaluation of compositional risks arising from interactions across modalities and time. The benchmark covers 11 non-singleton conditioning configurations with 11,024 attack instances, revealing systematic weaknesses in current multimodal safety guards. The findings highlight the need for enhanced compositional risk perception to safeguard AI systems, marking a sign