Computer Vision: January 2026 Week 5

1. Unifying Masked Diffusion Models with Various Generation Orders and Beyond

arXiv:2602.02112

Why Novel: First framework to unify masked diffusion models (MDMs) across arbitrary generation orders. Shows that generation quality depends critically on unmasking order, and introduces learnable ordering that automatically discovers optimal sequences.

Key Innovations:

• Order-expressive MDM (OeMDM) generalizes MDLM to support arbitrary unmasking orders via position-dependent noise schedules
• Learnable-order variant (LoMDM) discovers optimal generation orders from data
• 22 formal mathematical results grounding the framework
• Demonstrates that order choice dramatically affects sample quality — a previously underappreciated design axis

Evidence:

• sec:3 — Theoretical framework unifying various orderings in MDMs via position-dependent schedulers
• sec:4 — Learnable ordering mechanism and training procedure
• tab:1 — Comparison across different generation orders showing quality dependence
• tab:2 — LoMDM outperforming fixed-order baselines

Impact: Establishes generation order as a first-class design dimension for masked diffusion, enabling automatic discovery of optimal generation sequences.

2. Entropy-Based Dimension-Free Convergence and Loss-Adaptive Schedules for Diffusion Models

arXiv:2601.21943

Why Novel: Derives the first dimension-free convergence bound for diffusion models using Shannon entropy instead of KL divergence. The $O(H^2/K)$ bound removes exponential dimension dependence that plagued prior analyses.

Key Innovations:

• Dimension-free $O(H^2/K)$ KL divergence bound where $H$ is Shannon entropy of target
• Loss-adaptive noise schedules derived from the theoretical framework
• 26 formal mathematical results including convergence theorems
• Practical schedule recommendations that improve FID without retraining

Evidence:

• thm:1 — Main convergence theorem with dimension-free bound
• sec:3 — Problem setup and entropy-based analysis framework
• tab:1 — Comparison of loss-adaptive vs standard schedules

Impact: Resolves the curse of dimensionality in diffusion convergence theory and provides principled schedule design.

3. Lookahead Sample Reward Guidance for Test-Time Scaling of Diffusion Models

arXiv:2602.03211

Why Novel: First derivative-free, Taylor-approximation-free reward guidance method for diffusion models. LiDAR uses lookahead sampling to estimate rewards at intermediate denoising steps without backpropagation through the model.

Key Innovations:

• Lookahead sampling estimates future reward without gradient computation
• Compatible with any reward model — no differentiability requirement
• Test-time compute scaling: more compute → better alignment
• Outperforms gradient-based guidance methods while being simpler

Evidence:

• sec:3 — LiDAR framework and lookahead sampling mechanism
• sec:4 — Theoretical analysis of reward estimation quality
• tab:1 — Comparison with gradient-based and sampling-based guidance methods
• fig:1 — Test-time scaling curves showing compute-quality tradeoff

Impact: Opens test-time compute scaling for diffusion models analogous to LLM test-time scaling, without requiring differentiable rewards.

1. NativeTok: Native Visual Tokenization for Improved Image Generation

arXiv:2601.22837

Enforces causal dependencies in VQ tokenization with ordered token prediction, closing the gap between reconstruction and generation quality.

2. Progressive Checkerboards for Autoregressive Multiscale Image Generation

arXiv:2602.03811

Balanced progressive checkerboard ordering enables parallel AR image generation at multiple scales simultaneously.

3. Know Your Step: Faster and Better Alignment for Flow Matching Models via Step-aware Advantage

arXiv:2602.01591

TAFS-GRPO enables few-step flow matching generation with per-step reward feedback, reducing alignment latency.

4. PromptRL: Prompt Matters in RL for Flow-Based Image Generation

arXiv:2602.01382

Jointly trains language model with flow-based generator via RL, increasing exploration diversity and reducing prompt overfitting.

5. EventNeuS: 3D Mesh Reconstruction from a Single Event Camera

arXiv:2602.03847

Self-supervised dense 3D mesh reconstruction from monocular event streams via neural implicit surfaces.

6. Learning Physics-Grounded 4D Dynamics with Neural Gaussian Force Fields

arXiv:2602.00148

Unifies 3D Gaussian perception with ODE-based neural dynamics for interactive physics-grounded video prediction.