Computer Vision: February 2026 Week 7

1. Learn from Your Mistakes: Self-Correcting Masked Diffusion Models

arXiv:2602.11590

Why Novel: First principled framework enabling masked diffusion models to both decode AND correct their own errors. ProSeCo (Progressive Self-Correction) adds a simple cross-entropy corrector loss that trains the model to recover from its own mistakes, enabling inference-time self-correction loops.

Key Innovations:

• Joint training objective combining standard MDM loss with self-correction loss using tied weights
• Treats model outputs as corrupted data, training the same model to denoise its own errors
• Inference interleaves corrective steps with unmasking, enabling dynamic refinement
• Achieves 2-3x faster generation at equal quality, or up to 1.3x accuracy improvement

Evidence:

• sec:1 — Framework motivation showing error accumulation in parallel decoding
• sec:3 — Self-correcting objective with weight tying and argmax transformation
• fig:2 — Visual demonstration of ProSeCo recovering from collapsed generation
• alg:1 — Training algorithm with minimal modifications to standard MDM

Impact: Addresses fundamental limitation of parallel decoding in masked diffusion - once unmasked, tokens were fixed. Now models can dynamically revise, enabling quality-speed tradeoffs and inference-time scaling.

2. Continuous Diffusion Models Can Obey Formal Syntax

arXiv:2602.12468

Why Novel: First training-free guidance method for continuous diffusion language models that enforces discrete syntactic constraints (regular expressions). Diffinity analytically computes the probability that a latent state decodes to a valid sequence, using this gradient to steer sampling.

Key Innovations:

• Analytical guidance score computed from regex constraint without classifier training
• Targets conditional distribution given validity, not heuristic token masking
• Up to 70% validity on complex JSON-schema regexes while maintaining perplexity
• More reliable than autoregressive constrained decoding under finite token budgets

Evidence:

• sec:1 — Challenge of enforcing discrete constraints in continuous latent dynamics

Impact: Enables structured generation (JSON, math formats, regexes) in continuous diffusion models, previously only possible in discrete or autoregressive paradigms.

3. FAIL: Flow Matching Adversarial Imitation Learning for Image Generation

arXiv:2602.12155

Why Novel: Reframes post-training alignment of flow-based generators as adversarial imitation learning, eliminating need for preference pairs or reward modeling. FAIL minimizes policy-expert divergence through a discriminator without explicit rewards.

Key Innovations:

• FAIL-PD exploits differentiable ODE solvers for low-variance pathwise gradients
• FAIL-PG provides black-box alternative for discrete or constrained settings
• Achieves competitive performance with only 13K demonstrations on FLUX
• Acts as regularizer against reward hacking when combined with reward-based optimization

Evidence:

• fig:1 — FAIL framework overview and performance gains with limited data
• tab:1 — Benchmark comparison with preference optimization methods
• fig:2 — Convergence dynamics showing PD stability vs PG speed
• tab:3 — FAIL as regularizer preventing reward hacking

Impact: Provides data-efficient post-training for flow models without costly preference data, generalizing to discrete image and video generation.

1. ForeAct: Steering Your VLA with Efficient Visual Foresight Planning

arXiv:2602.12322

Visual foresight planning for VLAs - generates imagined future observations in 0.33s to guide step-by-step manipulation, achieving +40.9% success rate improvement over baselines.

2. Best of Both Worlds: Multimodal Reasoning and Generation via Unified Discrete Flow Matching

arXiv:2602.12221

UniDFlow unifies understanding and generation via task-specific LoRA adapters and reference-based multimodal preference alignment, achieving SOTA across 8 benchmarks.

3. Conversational Image Segmentation: Grounding Abstract Concepts with Scalable Supervision

arXiv:2602.13195

Grounds abstract, intent-driven concepts into pixel-accurate masks via conversational interface with scalable supervision.

4. DynaGuide: A Generalizable Dynamic Guidance Framework for Unsupervised Semantic Segmentation

arXiv:2602.13020

Dual-guidance unsupervised segmentation combining zero-shot global labels with local CNN refinement, improving mIoU 17.5% on BSD500.

5. TG-Field: Geometry-Aware Radiative Gaussian Fields for Tomographic Reconstruction

arXiv:2602.11705

Adapts 3D Gaussian Splatting for CT reconstruction with geometry-aware constraints, improving medical imaging quality.

6. Electrostatics-Inspired Surface Reconstruction (EISR)

arXiv:2602.11642

Novel surface reconstruction via Poisson's equation and Green's functions, representing shapes as superposition of Gaussian charges.