Anatomy of Attention: How New Research Is Rewiring Image Fidelity, Video Realism, and Prompt Obedience

TokenFlow++: Enforcing Temporal Consistency in Text-to-Video via Cross-Frame Token Propagation

Temporal flickering between frames remains one of the hardest unsolved problems in text-to-video — this paper introduces a cross-frame token propagation mechanism that enforces coherence without retraining. Practitioners working on long-form video generation should test this immediately as a drop-in consistency layer.

Prompt Obedience Scoring: A New Metric for Evaluating Text-to-Image Alignment Beyond CLIP

CLIP scores have long been the lazy default for measuring how well generated images match prompts, but this work argues they systematically miss spatial and relational reasoning failures. The proposed scoring system could become a new evaluation standard that finally catches the compositional errors users actually complain about.

DiffAttn: Disentangling Content and Style in Diffusion Models Through Dual Attention Streams

Separating content from style at the attention level — rather than through LoRA hacks or textual inversion — is a cleaner architectural approach that yields sharper style transfer with less identity bleed. This is the kind of foundational work that tends to get quietly absorbed into major model releases six months later.

GeomDiff: Geometry-Aware Diffusion for Physically Plausible Scene Generation

Most text-to-image models still generate physically implausible geometry — objects floating, shadows pointing wrong directions — and GeomDiff tackles this by conditioning the diffusion process on inferred depth and surface normals. For product visualization and architectural rendering use cases, this is a meaningful step toward images that don't require manual correction.

ControlVideo-3D: Lifting 2D Motion Controls Into 3D-Consistent Video Generation

Extending ControlNet-style pose and depth conditioning into 3D-consistent video is technically non-trivial, and this paper shows a working approach that maintains multi-view coherence across frames. Useful signal for anyone building avatar, virtual try-on, or synthetic training data pipelines.