AI video models and startups: every new video generation and editing model and tool (Sora, Veo, Runway, Kling, Luma, Pika, Seedance, Higgsfield, Midjourney…), open video models and research, avatar and dubbing tools, and the news of the companies building them. Newest first, every story told once.
Frametell is a local-first TypeScript project that turns your notes and material into explainer videos using an AI director agent, and works with any model. The repo gained 117 stars in its first week.
Kling AI showcased Kling 4.0, its generative AI video model aimed at enterprise production, at Advertising Week New York. A separate hands-on test examined how long Kling 4.0 keeps a product visually consistent.
afaqs! analysis says most AI video resizing quietly distorts product visuals, a practical caution for marketers using AI video tools.
ComfyUI is running a two-week contest with $25,000 in prizes, including RTX 5090s, for open-source rebuilds of Higgsfield creative apps on the Comfy Developer Platform.
The Western Weekender reviews Pollo AI for video creators, covering its agent, model options and control features.
Nerdbot published a comparison of AI video models Seedance 2.5, Veo 3, Wan 3.0 and MiniMax H3.
Quasa review covers Syllaby, an AI tool for producing faceless and avatar videos as a workflow for creators.
Quasa review covers Synthesys, an AI video agent offering multi-model orchestration and marketing video generation.
CutBay is a tool that pairs chat-based AI video generation with canvas editing, per famemagazine's coverage.
TaoMate H3 FL2VA is an image-to-video LoRA for the TaoMate H3 model; a developer converted it to BF16 ComfyUI format in roughly 1.24 GB and shared it on Hugging Face.
India's NCPCR has summoned Google over AI 'kiss and hug' apps on the Play Store that generate intimate videos without consent. The commission plans child safety guidelines by December.
Why it matters: ByteDance's video model refresh raises the bar in a fast-moving AI video generation race.
ByteDance released Seedance 2.5, its next-generation AI video generation model, which posts describe as leading 2026 content-creation workflows. No pricing or availability details were given.
Why it matters: An open consumer-GPU game video foundation model could broaden access to interactive world models.
Matrix-Game 4.0 is an open-source video foundation model from Kunlun Tech that generates playable game video on consumer GPUs, which the company calls the first such consumer-GPU game model.
Why it matters: It reduces wasted computation in streaming video generation, a key efficiency bottleneck for autoregressive video models.
Speculative Forcing is a method that lets few-step autoregressive video generators skip unnecessary refinement by only running extra denoising on tokens that still need it, cutting computation for streaming video generation.
Rushilo is a video editing tool controlled by talking to Claude Code or Codex. Listed on Product Hunt.
NVIDIA Broadcast 2.3 adds RTX Video 4K upscaling and 4x frame generation for webcams and prerecorded video files, per multiple reports.
Techloy reviews Gemini Notebook's AI-generated study videos, assessing whether the feature is worth using for learners.
Why it matters: A dedicated 3D-consistency benchmark gives a clearer way to judge whether video generators build coherent worlds.
Orbia is a new benchmark for systematically evaluating whether AI-generated videos depict a consistent 3D world, addressing a gap in existing geometric and memory-based evaluations.
A paper presenting DeltaStreamer, an agentic streaming video understanding framework that observes, remembers and investigates frames to retain evidence before future queries arrive, unlike compression-based methods.
Why it matters: Shows frontier video generation moving from short clips to multi-minute, voice-driven film workflows.
Higgsfield now makes AI videos longer than 10 minutes with GPT-6 Astra, sharper visuals via GPT-Image-2.5, and voice filmmaking with GPT-Live.
VideoJev is a family of System One models that turns video understanding and interaction into structured decision tasks, replacing autoregressive answer decoding in multimodal LLMs with a shared decision head.
FlashVSR+ is a video super-resolution method that adds hybrid attention and reward-yield distribution matching to one-step diffusion models, making high-resolution inference cheaper and better aligned with human preferences.
BGR explains how to use Nvidia RTX GPUs, likely via RTX Video, to improve the quality of older videos. A how-to, not a new release.
A report suggests OpenAI's Sora video model may see renewed momentum as the company revisits its video generation strategy.
A TypeScript tool that makes Claude Code render images and play videos directly in the terminal, including screenshots, Playwright recordings and synced before/after QA views. Works with Ghostty and kitty terminals.
Why it matters: Sora's revival effort indicates OpenAI's renewed consumer video push.
The Information details OpenAI's plan to revive its Sora video generation product. Specifics are behind a paywall.
An open-source ComfyUI node that upscales long videos with Nvidia RTX Video Super Resolution while disk-backed processing keeps RAM use low. Supports denoising, deblurring, H.264/HEVC NVENC encoding and audio preservation.
A paper introducing VEDJE, a video-efficient discriminative joint encoder that compresses features within sampled frames while caching representations separately, using feature-change prediction as an auxiliary training signal for retrieval.
Takes for Mac is a Mac app for recording talking-head videos, where Claude Code handles the editing. Launched on Product Hunt with discussion and a link.
SupoClip is an open-source AI clipper that cuts long videos into short viral clips. Launched on Product Hunt with a link and discussion.
An open-source JavaScript motion design studio that runs locally: Claude builds the video step by step while you approve each stage, using your own Claude Code or API key. Gained 526 stars in its first week.
Nikon disqualified the first-place Small World in Motion video by Dr. Ning Xu, saying it broke contest rules on generative AI.
Refs is a Product Hunt launch that gives Claude better reference material for working on videos. The source text gives few other details.
Pollo AI uses GPT-5.6, GPT-6 Astra and GPT-Image-2.5 to help creators turn ideas into detailed images and cinematic video ads.
Why it matters: Extends 3D world models from frozen scenes to animated, view-consistent cinemagraphs.
OuroWorld is a mask-free framework that animates any static 3D Gaussian Splatting scene into a seamlessly looping cinemagraph from any viewpoint. A vision-language model infers plausible dynamics and guides a video model to synthesize the motion.
Why it matters: Unifies grounded visual retrieval across images and video in one agent.
OneSearch-VL is a unified multimodal deep research agent for single-image, multi-image and video tasks. It centers on visually grounded retrieval, preserving dependencies between visual anchors, entity relations, source facts and answer operations.