Six ECCV 2026 Papers – Vision, Video Generation, Visual‑Language Navigation

ECCV (European Conference on Computer Vision) is one of the top three conferences in computer vision, alongside CVPR and ICCV. The 2026 edition received 10,473 valid submissions and accepted 2,883 papers, yielding an acceptance rate of about 27.5%. Gaode had six papers accepted.

Paper 01 – Geometry-Guided Reinforcement Learning for Multi-view Consistent 3D Scene Editing

Domain: Computer Vision / Generative AI / 3D Content Editing. The authors observe that while generating multi‑view consistent 3D content is difficult, verifying 3D consistency is relatively feasible, making reinforcement learning (RL) a viable solution. They propose RL3DEdit, a single‑forward‑pass framework driven by RL, and design a new reward signal derived from the 3D foundation model VGGT. VGGT provides a confidence map and camera pose error, which are used as rewards to anchor 2D editing priors onto a manifold with 3D consistency. Experiments show RL3DEdit achieves stable multi‑view consistency, superior editing quality compared with state‑of‑the‑art methods, and higher efficiency. The code and models are released.

Paper 02 – OmniDance: Multimodal Driven Dance Video Generation with Large‑scale Internet Data

Domain: Computer Vision / Video Generation / Digital Humans. Existing methods struggle to simultaneously achieve high visual quality and expressive full‑body dance motions due to limited large‑scale high‑quality dance video data and lack of a unified framework that integrates music. The authors introduce the CIPE‑Dance dataset (≈300k high‑quality clips, >400 h video) with text annotations describing choreography. OmniDance builds on a video generation backbone and adds music cross‑attention, a music‑text progressive specialization structure, a curriculum learning strategy from easy to hard, and a modality‑specific CFG inference strategy. This enables unified generation from text‑image, music‑image, and music‑text‑image inputs. Experiments demonstrate leading performance in visual quality, motion fidelity, and music/text alignment.

Paper 03 – CoMaTrack: Competitive Multi‑Agent Game‑Theoretic Tracking with Vision‑Language‑Action Models

Domain: Vision‑Language Navigation. Traditional embodied visual tracking relies on single‑agent imitation learning, which depends on expert demonstrations and generalizes poorly. Inspired by competitive evolution in nature, the authors propose CoMaTrack, a competitive multi‑agent game‑theoretic RL framework. It transforms static data‑driven imitation learning into adversarial interactive learning: a tracking agent and an adaptive opponent engage in a dynamic game across diverse environments, providing a self‑enhancing “arms race” training loop. They also release CoMaTrack‑Bench, the first benchmark for competitive embodied visual tracking with multiple scenes and language commands. Experiments show CoMaTrack achieves the best performance on both standard and adversarial benchmarks; a 3 B parameter vision‑language model trained with this framework surpasses prior 7 B single‑agent methods, reaching success rates of 92.1 % (simple tracking), 74.2 % (distractor tracking), and 57.5 % (adversarial tracking).

Paper 04 – ConceptWeaver: Weaving Disentangled Concepts with Flow

Domain: Computer Vision / Video Generation. While flow‑based generative models can produce high‑quality visual content, extracting and disentangling multiple visual concepts from a single reference image and recombining them in new scenes remains challenging. Existing personalization methods trade off concept fidelity against textual controllability and often rely on specific architecture modulation spaces. The authors first design a differential probing method to analyze how individual concept tokens affect the flow model’s velocity field, revealing three stages: Blueprint (structure building), Instantiation (strong, naturally disentangled concepts), and Refinement (texture polishing). Based on this, ConceptWeaver learns concept‑specific semantic offsets from a single reference image and injects them at the appropriate generation stage via stage‑aware optimization and ConceptWeaver Guidance. Experiments show high concept fidelity, controllable composition, and visual quality across single‑concept generation, multi‑concept composition, reference‑to‑video generation, motion concept transfer, and image editing.

Paper 05 – Towards High‑Resolution Visual Perception via Hierarchical Entity Exploration

Domain: Multimodal Agent / High‑Resolution Image Perception. High‑resolution image understanding requires multimodal large language models (MLLMs) to locate and exploit fine‑grained visual clues despite limited input resolution. Fixed‑resolution scaling causes geometric distortion and detail loss. Training‑based agents that select regions are costly and hard to transfer, while training‑free agents (e.g., ZoomEye, RAP) suffer from background fragmentation. The authors find that cropping benefits stem from background suppression rather than simple magnification. They propose Hierarchical Entity Exploration (HEE), a training‑free, model‑agnostic multimodal visual exploration agent. HEE uses a frozen detector to propose entity candidates, employs dual scoring to assess evidence sufficiency, and recursively clusters, merges, and explores around whole objects when needed, guided by confidence‑based backtracking. Experiments on Visual Probe, HR‑Bench, and MME‑RealWorld benchmarks show significant performance gains for open‑source MLLMs (Qwen2.5‑VL, LLaVA‑OneVision) and outperform training‑free baselines with lower compute cost.

Paper 06 – Towards Memory‑Efficient Autoregressive Video Generation via Instance‑Specific Parametric Absorption

Domain: Computer Vision / Video Generation. Autoregressive video generation relies on a growing Key‑Value (KV) cache, leading to linear memory growth and reduced throughput for long videos. Existing compression methods discard redundant KV tokens but can break long‑range dependencies, causing visual artifacts. The authors introduce Instance‑Specific Parametric Absorption (ISPA), which compresses the KV cache by “absorbing” history rather than cutting it. During a warm‑up phase, ISPA collects global and local attention outputs, models their difference as a reconstruction problem, and solves for instance‑specific weight modulation via closed‑form least squares. This converts some global attention layers to local ones and compensates removed context with updated projection weights. An online softmax and LSE‑based decomposable attention mechanism enables low‑cost signal collection. Experiments across text‑to‑video and speech‑to‑video tasks on models ranging from 1.3 B to 14 B parameters remove up to ~50 % of KV cache with near‑lossless visual quality; combined with W8A8 quantization, inference speed improves by 1.86×.

For more detailed analyses, stay tuned to future Gaode technology updates.