Searcharxiv⌕ Search
Home/Archive/arXiv

arXiv · 2610.37626

Know, then Draw: A New Learning Paradigm for Visual Generation

Yucheng Mao·Chongjian Ge·Zhuowen Tu·Hao Tan·Hanwen Jiang
View original source ↗Read source document ↗Export metadata ↓

Abstract

Visual generation requires both high-level semantic knowledge and fine-grained pixel details. We propose a learning paradigm that assigns these roles to autoregressive pretraining for knowledge acquisition and diffusion post-training for high-fidelity rendering. Within this framework, we study the objective, representation, and scaling of knowledge acquisition and rendering. Controlled comparisons identify autoregression as the stronger pretraining objective for downstream rendering, improving both learning efficiency and final generation quality. We further find that AR-only generation quality does not necessarily correlate with final performance; semantic quality matters more. This motivates a purely semantic tokenizer trained with language supervision alone, without an image decoder or reconstruction loss. We establish compute-optimal scaling laws for knowledge acquisition and rendering, and find that autoregressive pretraining lowers the fitted asymptotic rendering loss compared with pixel diffusion models trained from scratch. Guided by the laws, we train a 3B-parameter text-to-image model with only 6,600 H100 GPU-hours, achieving 86 on GenEval and 87 on DPG-Bench and validating the proposed paradigm at scale.

Explore related subjects

cs.CV

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Yucheng Mao, Chongjian Ge, Zhuowen Tu, Hao Tan, Hanwen Jiang. 2026-10-27. Know, then Draw: A New Learning Paradigm for Visual Generation. https://arxiv.org/abs/2610.37626

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

DyRA: Portable Dynamic Resolution Adjustment Network for Existing Detectors

Achieving constant accuracy in object detection is challenging due to the inherent variability of object sizes. One effective approach to this problem involves optimizing input resolution, referred to as a multi-resolution strategy. Previous approaches to resolution optimization have often been based on pre-defined resolutions with manual selection. However, there is a lack of study on run-time resolution optimization for existing architectures. This paper introduces DyRA, a dynamic resolution adjustment network providing an image-specific scale factor for existing detectors. This network is co-trained with detectors utilizing specially designed loss functions, namely ParetoScaleLoss and BalanceLoss. ParetoScaleLoss determines an adaptive scale factor for robustness, while BalanceLoss optimizes overall scale factors according to the localization performance of the detector. The loss function is devised to minimize the accuracy drop across contrasting objectives of different-sized objects for scaling. Our proposed network can improve accuracy across various models, including RetinaNet, Faster-RCNN, FCOS, DINO, and H-Deformable-DETR. The code is available at https://github.com/DaEunFullGrace/DyRA.git.

cs.CV↗
arXiv2026

Smart Surveillance: Deep Learning and Hybrid Approaches for Dynamic Scene Analysis, Object Detection and Motion Tracking

Event-triggered recording reduces the footage a surveillance camera must store, but the trigger decides how much idle video is kept, how many events are missed, and how much computation is spent. We compare six motion triggers (frame differencing, dense optical flow, MOG2, KNN, CNT, GSOC), a CNN detector, strided detection, and three motion-gated hybrids under one recording policy, on a benchmark of seven static-camera scenes (5,800 frames, 27 annotated events) with thresholds transferred to unseen scenes by leave-one-video-out. Detector-based triggers stored 19-28% of idle frames, against 36-89% for motion triggers. Under simulated camera jitter, frame differencing, optical flow, MOG2 and KNN stored 97-100% of idle frames, while detector triggers stayed below 29%. Running the detector on every 8th frame kept 99.7% of event frames at 9.8 ms of CPU per frame, cheaper and more robust than gating it with background subtraction. On VisDrone, raising YOLO11n's input from 640 to 1280 px raised AP from 15.5 to 26.8, more than a larger model, and INT8 quantization cut latency by 17% for 0.3 AP. In tracking, longer track memory raised IDF1 from 70.5 to 76.7, but every tracker recovered only 39-44% of overlapped people, because detection rather than association failed.

cs.CV↗
arXiv2026

An Embarrassingly Simple Data Selection Strategy for Million-Scale Multi-Modal Instruction-Tuning

Recent research demonstrates that pretrained large language models (LLMs) require minimal but carefully curated data for instruction-tuning. Building on LLMs, multi-modal models further contend with far larger data volumes, heterogeneous sources, and cross-modal alignment, amplifying both the necessity and the complexity of data selection. However, existing methods rely on deeply extracted embeddings, gradients or narrow quality scores that are time-consuming to compute and limited in expressiveness, frequently failing to surpass a simple random-sampling baseline. To overcome these limitations, we step back from complex embedding spaces and target the fundamental objective of instruction tuning: how and what knowledge should be elicited from the pretrained model and exposed to users. We leverage interaction \underline{S}tyles as a proxy for structural diversity, and decompose data quality into capability-specific \underline{S}cores, yielding an explicit yet \underline{R}ich (SSR) profile for each sample. We further cast the selection as submodular coverage maximization over the capability--style grid, providing an approximation guarantee. SSR consistently delivers superior performance across two distinct architecture stacks, LLaVA-1.5 (Vicuna + CLIP) and LLaVA-OneVision (Qwen2 + SigLIP), two large-scale data pools, six data budget settings, and 14 multi-modal benchmarks. To our knowledge, SSR is the first data selection method validated at million-scale, achieving 99.1% of full-data performance using only 30% of a 2.6M pool. Furthermore, the SSR profile models and selected subsets transfer training-free across data pools and model architectures with capability extensibility.

cs.CV↗