#video-generation

The discussion moved quickly from sample quality to the larger claim: one multimodal backbone for visual generation and action prediction.

The discussion split between scientific utility and obvious misuse concerns. NEvo uses a digital twin of the visual brain as a reward model, then evolves AI-generated clips to maximize predicted activation in target regions.

Meta has rolled out Muse Image across Meta AI, meta.ai, Instagram Stories in the U.S., and WhatsApp in limited countries. The notable shift is an image model that uses search, code execution, self-refinement, and a watermarking system inside consumer social products.

NVIDIA Research’s MOTIVE targets a specific video-model bottleneck: which fine-tuning clips actually improve motion. The ICML 2026 honored paper reports a 74.1% human preference result against the base model.

Meituan’s LongCat team released an audio-driven avatar video model with Diffusers examples and an MIT license on Hugging Face. The project compares against HeyGen, Kling Avatar 2.0, and OmniHuman-1.5.

At Google I/O 2026, Google DeepMind unveiled Gemini Omni — its first model capable of generating video from any input including text, images, audio, and video. Combining Gemini's intelligence with Google's generative media systems, it is available now through the Gemini app and YouTube Shorts.

Google revealed Gemini Omni at I/O 2026—a "world model" that processes text, audio, images, and video together to simulate physical environments. Unlike Sora or Runway, it lets users edit footage through natural language and maintains scene consistency across modifications. It replaces Veo in the Gemini app immediately.

NVIDIA Labs released SANA-WM, a 2.6B parameter open-source world model capable of generating up to one minute of 720p video. The relatively small model size and open-source availability make it a significant contribution to accessible video generation research.

A video believed to be from Google's unreleased 'Omni' video generation model has leaked, drawing 1,300+ upvotes on r/singularity. Users particularly noted the model's unusually coherent text rendering - a persistent weakness in current video generation models.

Google introduced Veo 3.1 Lite as its most cost-effective video generation model, priced at less than 50% of Veo 3.1 Fast while keeping the same speed. The model is rolling out through the paid tier of the Gemini API and Google AI Studio, broadening access to higher-volume video app use cases.

Netflix’s VOID reached Reddit as an open research release aimed at removing objects from video and repairing the interactions those objects caused in the scene. The notable details are the CogVideoX base, a two-pass pipeline, Gemini+SAM2 mask generation, and a 40GB+ VRAM requirement.

Together AI said on April 3, 2026 that Wan 2.7 from Alibaba Cloud is now available on its platform. The accompanying product post says text-to-video is live now, with image-to-video, reference-to-video, and video edit workflows rolling out on the same API, auth, and billing surface.