Text to Video News and AI Updates

Latest Text to Video news, research, benchmarks, product releases, and industry adoption updates.

Latest Articles

• SemComp-Bench: Benchmarking Semantic Task Completion in Video Generation — SemComp-Bench introduces outcome-oriented evaluation for video generation — measuring whether models complete intended tasks with semantic fidelity, not just visual realism. Experiments show current models consistently fail to achieve both task goals and semantic grounding simultaneously.
• Adding Voice Cloning to Text-to-Audio-Video Models with a Single Zero-Initialised Layer — A single zero-initialized linear layer added to a 5B text-to-audio-video model enables state-of-the-art voice cloning across 30 speakers — outperforming five TTS baselines on speaker similarity — and the audio-only path runs approximately 30x faster than full audio-video diffusion.
• CaliBench: Are the Stochastic Dynamics of Video World Models Physically Calibrated? — CaliBench tests whether video generation models reproduce physically correct stochastic outcomes: across six models (Veo 3.1, Runway Gen-4.5, WAN-2.7, SeeDance-2.0, HappyHorse-1.0, Cosmos3-Super), most are significantly miscalibrated — with Veo 3.1 collapsing to a single dice face rather than generating uniform outcomes.
• SCOPE: Score-Isolated Agentic Optimization for Video World Models — SCOPE introduces auditable inference-time adaptation of frozen video world models by formalizing updates as bounded typed state transitions, achieving +14.24 Physics-IQ benchmark improvement while revealing that inference-time gains do not transfer uniformly across backbone architectures.
• Equilibrium Forcing: Adaptive Video Generation Without Noise Conditioning — Equilibrium Forcing (EqF) removes noise-level conditioning from video diffusion and flow-matching models, decoupling training from sampling and enabling inference-time algorithms that adapt to feedback from the generated sample — improving video quality and temporal consistency on autoregressive video generation benchmarks.
• MLLM-Guided Semantic Correction for Text-to-Video Generation — A training-free framework injects MLLM feedback directly into the diffusion sampling loop during video generation, catching and correcting semantic errors — missing objects, wrong attributes, mismatched actions — mid-generation without modifying model weights or restarting the process.
• Rethinking Token-wise Feature Caching: Accelerating Diffusion Transformers with Dual Feature Caching — DuCa challenges the assumption that 'important tokens' must always be computed in diffusion transformers—finding that consistent computation of selected tokens is unnecessary and sometimes worse than random selection—then introduces alternating aggressive/conservative caching with random token selection, published in IEEE Transactions on Image Processing.
• SemVideo: Reconstructs What You Watch from Brain Activity via Hierarchical Semantic Guidance — SemVideo reconstructs video clips from fMRI brain activity using a three-tier semantic hierarchy — static anchor descriptions, motion narratives, and holistic summaries — achieving state-of-the-art temporal coherence and semantic alignment on the CC2017 and HCP brain imaging datasets.
• FrescoDiffusion: 4K Image-to-Video with Prior-Regularized Tiled Diffusion — FrescoDiffusion generates globally coherent 4K video from a single image without training, by using a low-resolution latent video trajectory as a global prior that guides tiled high-resolution denoising via a weighted least-squares fusion at every diffusion step.
• ForgeWM: Progressive Causal Training for Few-Step Action-Conditioned Video World Models — ForgeWM converts a bidirectional video generator into 1-, 2-, and 4-step interactive world models for games via progressive causal distillation, achieving the lowest LPIPS and top action-sign accuracy on Minecraft — while a replay-time refinement mode matches 4-step reference quality.