Computer Vision: January 2026 Week 1

1. Guiding Token-Sparse Diffusion Models

arXiv:2601.01608

Why Novel: Discovers that classifier-free guidance (CFG) fails for token-sparse diffusion models and introduces Sparse Guidance (SG), a finetune-free, test-time method that uses two sparsity levels to create a capacity gap for effective guidance — a fundamentally new guidance paradigm.

Key Innovations:

• Identifies that CFG provides limited benefits for sparsely-trained diffusion models, a previously unrecognized failure mode
• SG uses differential sparsity levels rather than conditional/unconditional gap, requiring no additional training
• Achieves 1.58 FID on ImageNet-256 benchmark, state-of-the-art for efficient diffusion models
• Reduces compute (GFLOPs) while improving both quality and diversity simultaneously

Evidence:

• fig:2 — Demonstrates CFG failure mode for token-sparse models — limited quality improvement with increasing guidance scale
• tab:1 — SG improves upon CFG across masking and routing sparsity strategies
• tab:2 — SG outperforms other guidance methods by significant margins in FID and GFLOPs
• tab:3 — Achieves 1.58 FID on ImageNet-256, competitive with dense models at fraction of compute

Impact: Unlocks practical deployment of efficient sparse diffusion models by solving the guidance problem, enabling high-quality generation at significantly reduced compute.

2. Towards Agnostic and Holistic Universal Image Segmentation with Bit Diffusion

arXiv:2601.02881

Why Novel: First framework to make universal image segmentation truly vocabulary-agnostic by formulating segmentation as discrete diffusion over bit representations, removing dependence on predefined label sets and enabling holistic scene understanding.

Key Innovations:

• Adapts discrete diffusion to segmentation via analog bit diffusion with location-aware palette (LAP) using 2D gray-code layout
• Input-scaled noise schedule and $\tanh$ activation with $x$-prediction and sigmoid loss for stable training
• Vocabulary-agnostic: works across semantic, instance, and panoptic segmentation without label-specific heads
• Competitive with SOTA specialized models while being a single unified architecture

Evidence:

• fig:1 — Overview showing modifications to base diffusion model and their cumulative performance gains
• tab:1 — Performance comparison across different encoding types validating bit diffusion approach
• tab:2 — Competitive results against SOTA specialized segmentation models on public validation set
• tab:4 — Ablation results validating each component contribution

Impact: Opens a new paradigm for universal segmentation where a single diffusion model handles all segmentation tasks without vocabulary constraints, simplifying deployment and enabling open-world scene understanding.

3. Understanding Reward Hacking in Text-to-Image Reinforcement Learning

arXiv:2601.03468

Why Novel: First systematic analysis revealing that reward hacking in text-to-image RL follows universal artifact patterns across different reward types (aesthetic, consistency, human preference), providing both diagnostic tools and understanding of failure modes.

Key Innovations:

• Demonstrates that all common T2I reward types (aesthetic, prompt-image consistency, human preference) are exploitable via artifact generation
• Reveals universal pattern: models learn to generate specific visual artifacts that reliably inflate reward scores
• Creates curated artifact diagnostic dataset for evaluating reward model robustness
• Tests accuracy of reward models in distinguishing artifact-free vs artifact-laden images

Evidence:

• fig:1 — Evolution of metrics over training showing reward score inflation while quality degrades
• tab:2 — Accuracy of different reward models in assigning higher scores to artifact-free images — revealing widespread vulnerability
• fig:2 — Visual examples showing artifact generation patterns across different reward training setups

Impact: Essential reading for anyone applying RL to image generation — demonstrates that naive reward optimization is fundamentally broken and provides diagnostic framework for building more robust training pipelines.

1. Improving Flexible Image Tokenizers for Autoregressive Image Generation

arXiv:2601.01535

Identifies and solves the generation bottleneck in flexible 1D image tokenizers via ReTok with redundant token padding and hierarchical semantic regularization.

2. NextFlow: Unified Sequential Modeling Activates Multimodal Understanding and Generation

arXiv:2601.02204

Unified decoder-only Transformer trained on 6T interleaved text-image tokens achieving both understanding and generation via next-scale prediction for visuals.

3. VINO: A Unified Visual Generator with Interleaved OmniModal Context

arXiv:2601.02358

Unifies image and video generation and editing under a single diffusion backbone conditioned on interleaved omni-modal context with token-boundary identity preservation.

4. ActErase: A Training-Free Paradigm for Precise Concept Erasure via Activation Patching

arXiv:2601.00267

Training-free concept erasure in diffusion models via FFN activation patching, achieving SOTA erasure across nudity, artistic style, and object categories.

5. Multi-Subspace Multi-Modal Modeling for Diffusion Models: Estimation, Convergence and Mixture of Experts

arXiv:2601.01475

Provides theoretical foundation for why diffusion models learn from small datasets via multi-subspace multi-modal manifold modeling with 36 formal results.

6. Pixel-to-4D: Camera-Controlled Image-to-Video Generation with Dynamic 3D Gaussians

arXiv:2601.00678

Per-pixel Gaussian splats with linear/angular velocities and accelerations enabling camera-controlled image-to-video generation from a single image.

7. AdaGaR: Adaptive Gabor Representation for Dynamic Scene Reconstruction

arXiv:2601.00796

Replaces Gaussian primitives with Adaptive Gabor Primitives that learn frequency weights for better high-frequency texture capture in dynamic scene reconstruction.