Computer Vision: March 2026 Week 10

1. ZipMap: Linear-Time Stateful 3D Reconstruction with Test-Time Training

arXiv:2603.04385

Why Novel: Prior state-of-the-art feedforward 3D reconstructors (VGGT, Fast3R, $\pi^3$) all use full attention with O(N²) cost, making them impractical beyond ~100 frames. ZipMap is the first to achieve competitive quality with true linear-time inference by replacing global attention with local windows and a TTT fast-weight memory that accumulates scene state.

Key Innovations:

• [object Object]
• [object Object]
• [object Object]

Evidence:

• fig:2 — undefined
• tab:1 — undefined
• tab:2 — undefined
• tab:3 — undefined

Impact: Makes reconstruction from long monocular video (100s–1000s of frames) tractable in feedforward models, removing the primary scalability bottleneck in neural visual geometry.

2. Order Is Not Layout: Order-to-Space Bias in Image Generation

arXiv:2603.03714

Why Novel: While prompt sensitivity in T2I has been studied, this is the first paper to isolate and formally benchmark a positional/sequential inductive bias inherited from language model pretraining — where first-mentioned entities default to the left and first-mentioned roles default to dominant subjects, even when image or spatial context says otherwise.

Key Innovations:

• [object Object]
• [object Object]

Evidence:

• fig:1 — undefined
• fig:2 — undefined
• tab:3 — undefined

Impact: Forces rethinking of prompt engineering, compositional generation evaluation, and training data curation — any spatial-reasoning task using T2I must account for this systematic language-order confound.

3. Generalization Properties of Score-matching Diffusion Models for Intrinsically Low-dimensional Data

arXiv:2603.03700

Why Novel: Prior theoretical analyses gave pessimistic convergence rates scaling with the ambient dimension $D$, which contradicted the strong empirical generalization of diffusion models on high-dimensional image data. This paper provides the first formal proofs that depend only on $d$, using covering/packing number arguments on the data manifold.

Key Innovations:

• [object Object]
• [object Object]

Evidence:

• def:1 — undefined
• def:2 — undefined
• fig:1 — undefined

Impact: Provides the theoretical foundation explaining why diffusion models generalize far better than ambient-dimension theory predicts, guiding future architecture and data-efficiency research.

1. MERG3R: A Divide-and-Conquer Approach to Large-Scale Neural Visual Geometry

arXiv:2603.02351

Reconstructs camera poses and geometry from 1,000+ unordered images via hierarchical partitioning + merging of VGGT-style sub-reconstructions, achieving state-of-the-art accuracy at memory-scalable cost.

2. LoGeR: Long-Context Geometric Reconstruction with Hybrid Memory

arXiv:2603.03269

Hybrid memory module combining lossless local attention with O(N) TTT-style global memory enables feedforward geometric reconstruction on minute-long videos without quadratic blowup.

3. DREAM: Where Visual Understanding Meets Text-to-Image Generation

arXiv:2603.02667

Unified model trained jointly on visual understanding and T2I generation via SD-encoder continuous tokens + semantically aligned decoding; outperforms specialist models on both tasks when trained on CC12M.

4. Phys4D: Fine-Grained Physics-Consistent 4D Modeling from Video Diffusion

arXiv:2603.03485

Integrates physics simulation curriculum into video diffusion training, achieving consistent rigid and deformable dynamics that vanilla video diffusion models fail to reproduce.

5. A Simple Baseline for Unifying Understanding, Generation, and Editing via Vanilla Next-token Prediction

arXiv:2603.04980

Wallaroo shows that decoupled visual encoding pathways with standard next-token prediction on Qwen2.5-VL matches specialized unified models on understanding benchmarks while also generating and editing images.