Tag
Pika introduces H3 Flex, a post-trained MiniMax H3 text-to-video model that generates in six steps (vs. 50), supports up to 12 reference files, exports five audio stems, finishes at up to 4K, and is released with open weights — ranking #1 among open-weights models on DesignArena's leaderboard.
Pika H3 Flex 是一款新发布的开源权重视频生成模型，主打专业迭代工作流，可在 7 秒内生成 15 秒带音轨的视频，支持导出音频分轨、最高 4K 分辨率草稿模式，以及对角色、声音和运镜的后期编辑。
Pika 发布开源权重的视频生成模型 Pika H3 Flex，支持六步采样生成带同步音频的视频，最多可输入 12 个参考，本地推理支持 480p/768p，通过 Pika API 可扩展至 4K 与音频分轨。
Kling 4.0 has just been released, with claims that its AI-generated video output is nearly indistinguishable from reality, signaling a major leap for AI video generation and potential disruption to traditional filmmaking.
A single prompt had Claude Fable 5.5 generate a 3-minute original animated film, covering the progress of human civilization through the next 30 years, with fully original visuals and soundtrack and no stock assets used.
Toyota Research与宾大提出Temporal Specification Guidance (TSG)，一种免训练的推理时框架，将时序逻辑规格转化为可微目标，通过梯度引导注意力偏置来改善文本到视频模型对顺序、时序与持续性约束的遵循，在五个模型上比CFG提升9.5–20个百分点的规格满足率。
提出 Routing Policy Optimization（RPO），在扩散生成器内部通过随机注意力路由引入探索，以对齐奖励信号，同时保留原始确定性采样器，并在文本到图像/视频生成上实现相比 Flow-GRPO 方法近 4 倍的训练效率提升。
StreamErase 实现了亚秒级延迟的实时流式视频物体擦除，通过滑动窗口因果注意力与注意力蒸馏将双向模型的擦除能力迁移到单向流式生成器，在单 GPU 上可达 26 FPS，并提出 LongErase 基准。
Grok Imagine 宣布 Video 1.5 Lite 模型已在其 API 上线，作为文本生成视频和图像生成视频的新主力模型，定价为 480p 每秒 0.02 美元、720p 每秒 0.03 美元、1080p 每秒 0.14 美元。
美团开源了13.6B参数的LongCat-Video视频生成模型，原生支持文本到视频、图像到视频和视频续写，可在720p/30fps下生成数分钟长视频且无色彩衰减或质量下降，权重采用MIT许可证并允许商业使用。
Creator Foyege (@Magncsans) spent ten days hand-drawing over 300 storyboards and burning through 100,000 credits to produce a Cthulhu-inspired sci-fi short film called Onyx, inspired by Love, Death & Robots, with 80% of the visuals assisted by AI (white-model) generation.
A tweet hypes the upcoming Fable 5.5 video model, claiming its launch video was generated from a single prompt and that some users are already in a gray rollout, positioning it against Opus 5.5, Remotion, and Hyperframe.
VTR-Bench introduces a systematic benchmark for evaluating visual text rendering in video generation models, covering 300 prompts across five scenario categories, an automated evaluation pipeline with human alignment, and a Keyframe-Guided Agentic Framework. Experiments on 11 state-of-the-art models reveal widespread difficulties in rendering scene text, with the best model achieving a word error rate of 0.250.
Utopai Studios unveiled PAI and Utopai X, which ranks as the #2 text-to-video model with audio on Artificial Analysis' global leaderboard and the highest-ranked from a U.S.-based company, with production-quality output used for sports, film, and TV projects.
The tweet shows a self-portrait video generated by Claude Opus 5.5, with effects described as comparable to a blockbuster movie, showcasing the powerful capabilities of AI models.
The article appears to be inquiring about the status or release date of the Muse AI model, possibly referring to NVIDIA's text-to-video technology.
WanPE is a 397B-parameter prompt enhancement model that improves cinematic planning in text-to-video generation, showing significant human preference boosts over raw prompts and competitive performance with commercial offerings.
The Biological Computing Co. partners with AWS to commercialize a neuron-derived AI video model that optimizes text-to-video generation using biological data, offering faster and cheaper inference on standard hardware.
Syllaby AI Avatar 2.0 lets users turn a script or idea into a complete presenter-led video featuring a realistic AI avatar, natural voice, B-roll, and subtitles—no actors or production crew required.
The H3 Max Director AI model can generate interactive scenes and plots from text inputs, enabling applications in short dramas and real-time games. This advancement highlights AI's potential in dynamic content creation.