Focused on computer vision, image processing, color science, and AI algorithms; sharing hardcore tech, engineering practice, and deep insights as a diligent AI technology practitioner.
Vision‑R1 introduces a 7B multimodal large language model that leverages 200K unsupervised CoT data, Modality Bridging, and Progressive Thinking Suppression Training to overcome data scarcity and over‑thinking, achieving 73.5% accuracy on MathVista—within 0.4% of OpenAI’s O1.
The AAAI 2026 paper by Tsinghua’s Huang‑Gao team shows that modeling Vision‑Transformer attention as a Block‑Circulant matrix and computing it with FFT reduces the quadratic complexity to O(N log N), delivering up to seven‑fold real‑world speedups without sacrificing accuracy.
InternVL-U, a 4‑billion‑parameter unified multimodal model released as open source, combines a 2B MLLM backbone with a 1.7B visual generation head and, through a reasoning‑centric data pipeline and Chain‑of‑Thought guidance, achieves superior understanding, generation, and editing performance that surpasses much larger 14‑20B models on multiple benchmarks.
Beyond-Hungarian replaces the costly Hungarian assignment in DETR with a differentiable, query‑free matching scheme that halves training latency, boosts large‑object AP by 4.2 points, and introduces a GT‑Probe module and dual‑loss framework, while detailing trade‑offs, ablations, and future challenges.
ArtiMuse, a new image aesthetic model unveiled at CVPR 2026 by Shanghai AI Lab and the China Academy of Art, combines a massive 10K fine‑grained dataset, a Token‑As‑Score scoring scheme, and unified textual‑and‑numeric feedback to deliver culturally aware, expert‑level art analysis and robust quantitative ratings.
Mind-Brush introduces a ‘think‑research‑create’ agentic framework that unifies intent analysis, multimodal evidence retrieval, and knowledge‑driven reasoning to transform text‑to‑image generation from static decoding into an active cognitive workflow, achieving large accuracy gains on the new Mind‑Bench benchmark and surpassing existing SOTA models.
The paper introduces BeautyGRPO, a reinforcement‑learning framework that combines a fine‑grained preference dataset (FRPref‑10K) with Dynamic Path Guidance to balance aesthetic enhancement and high‑fidelity preservation in portrait retouching, achieving superior metrics and user preference over existing SFT and RL models.
The article dissects the query‑activation imbalance in DETR‑based detectors, explains PaQ‑DETR’s pattern‑sharing and quality‑aware assignment mechanisms, and shows how these jointly raise detection mAP by up to 4.2% on COCO with less than 5% extra FLOPs.
The paper introduces MIGM-Shortcut, a self‑supervised method that learns controlled latent‑state dynamics to bypass redundant bidirectional attention in Masked Image Generation Models, achieving over 4× speed‑up on state‑of‑the‑art multimodal diffusion models like Lumina‑DiMOO while preserving image quality.
The article dissects EFSI‑DETR, a UAV small‑object detector that combines simulated frequency processing with dynamic semantic enhancement to overcome pixel scarcity, static fusion, and ignored frequency cues, delivering 188 FPS and a 5.8% APₛ gain on VisDrone while remaining lightweight.
