Video Generation

Today's Stories

1. 1

   Black Forest Labs releases Flux 3 with native audio video generation

   German AI company Black Forest Labs released Flux 3, a multimodal foundation model that learns from images, video, and audio together. The model can generate videos with native audio up to 20 seconds long for the first time, supporting text-to-video, image-to-video, video-to-video, keyframe-based transitions, multilingual dialogue, and agent-driven links between clips. In early evaluations using 10-second clips at 720p, Flux 3 was preferred over multiple rivals: 93 percent over Luma Ray 3.2, 77 percent over Runway Gen-4.5, and 69 percent over Grok Imagine Video. Against stronger competitors, it matched or came close to Seedance 2.0 and Gemini Omni Flash (each at 52 percent preference), which already serve major production workflows. BFL also developed Flux-mimic, a video-action model being tested on production tasks at Audi, suggesting the architecture extends to robotics applications.

   Flux 3 Image is set to launch in early access within the next few weeks, with improvements to complex prompts and multilingual text rendering. Action prediction will initially be offered through select partners. BFL plans open-weight access to the multimodal backbone under the name 'Flux 3 Dev,' with longer-term work on a single model combining perception, action, and language prediction.

2. 2

   Creator makes feature-length film using LLM, calls it a failure

   A content creator used an LLM (Claude Fable 5) to produce a feature-length movie adaptation of William Hope Hodgson's The House on the Borderland, which they uploaded to their YouTube channel alongside music videos and AI-generated albums. The experiment tests whether large language models can sustain creative output across a full-length project without human intervention—a question that bears on what kinds of creative work remain beyond current AI capabilities, even when given time and permission to operate independently.

   The creator frames the result as a failure despite releasing it, noting they judged it only by YouTube channel standards rather than film as a whole category; they attribute the shortcoming not to the LLM's agency or planning, but to other factors they discuss in the full piece.

3. 3

   Black Forest Labs launches FLUX 3 for image, video, and audio generation

   Black Forest Labs released FLUX 3, a multimodal AI model that generates images and combined audio/video clips up to 20 seconds from a single prompt. The model is jointly trained across image, video, and audio rather than combining separate models. It also extends to robotic vision and actions. FLUX 3 represents the company's first public video generation model and frames creative generation, simulation, computer use, and robotics as connected applications of a single capability. Black Forest Labs positions this as "visual intelligence"—models that can perceive, predict, and act across physical and digital environments—rather than treating each modality as a separate tool.

   FLUX 3 will be offered through four product lines: FLUX 3 Video, FLUX 3 Image, and FLUX 3 Act. The release is limited to start, meaning broader availability or pricing details may follow.

4. 4

   Runway launches Media Router to orchestrate generative AI models

   Runway launched the Media Router through its Runway Dev platform on Thursday, a tool that automatically selects the best image, video, or audio generation model for a developer request based on priorities like quality, speed, or cost. Runway says this is the first model router built specifically for generative media. Generative media models have exploded in number, making it difficult for developers to evaluate new releases and understand which model works best for each task. The router lets developers—including Adobe, Cloudflare, ElevenLabs, Expedia, Shutterstock, and Quora—integrate media generation directly into their products via API, while Runway positions itself as an orchestration layer rather than betting on a single model staying ahead.

   Developers can set preferences for the router's selection logic, including token pricing (a major cost concern in 2026 for enterprises using agentic AI) and model origin—for example, prioritizing American providers over Chinese models as the Trump administration explores potential bans on Chinese open AI models.

5. 5

   Synthesia launches AI coaching tool for sales, leadership training

   British AI startup Synthesia launched Roleplay Sessions on Wednesday, an interactive training product where employees practice high-stakes conversations—sales pitches, performance reviews, customer complaints—with an AI avatar that responds, challenges them, and scores their performance against a rubric. The company has already secured early customers including one of Europe's top three companies by market cap, one of the top five Fortune 100, and one of the world's biggest recruitment companies. Synthesia is shifting from video generation to proving training actually changes behavior. Most corporate training informs and demonstrates but stops short of driving real learning; Synthesia argues the breakthrough comes from practice plus feedback. By layering rubrics, performance data, and analytics on top of its avatar technology, the company is positioning itself as a performance-management platform rather than just an AI avatar vendor—a more defensible business model. CEO Victor Riparbelli notes that when an entire sales team practices with an AI role player, managers can gather granular data on overall sales-force performance.

   Roleplay is currently an enterprise offering, but Synthesia plans to expand it to small businesses, prosumers, and potentially educational institutions in the next few months as inference costs continue to fall. The platform is also the first release under a broader "Sessions" framework the company plans to extend into job interviews and candidate screening.

6. 6

   AI Movies Made Simple: How One Developer Built a Feature Film Using Claude and Video Models

   A developer used Claude (an AI that understands and generates text) paired with Seedance 2.0 (a video generation model) to create a complete short film. The work involved fifteen iterations and a multi-stage pipeline: starting with a two-page story prompt, evolving it through Claude brainstorming sessions, generating reference images for characters and locations, and then producing video clips for each scene using AI-generated prompts. This demonstrates that non-specialists can now produce movie-quality content by combining existing AI tools without needing film crews or expensive equipment. The author showed that Claude Fable (a newer model variant) and Claude Opus (an older variant) produced comparable directing quality when tested head-to-head, meaning the real breakthrough is the accessible toolchain itself rather than a single breakthrough model.

   The final film is available on YouTube (https://www.youtube.com/watch?v=q40M08SOhGs). The pipeline uses Higgsfield as an API wrapper for Seedance, nano banana for reference image generation, and Claude Code for scriptwriting—a stack that appears reproducible by others interested in AI filmmaking.