FLUX 3 Pushes Past Image Generation Into Video, Audio, and Action
Original: Flux 3 View original →
Black Forest Labs framed FLUX 3 as more than another image model. The company describes it as a multimodal flow model that learns from images, video, and audio together, with the goal of becoming a shared backbone for visual intelligence. That is what made the Hacker News discussion interesting: the claim is not only better media generation, but a broader foundation that can also support action prediction.
The launch post says FLUX 3 can mix modalities and generate image, video, and audio outputs from text prompts or from visual references. For video, BFL says the model can create diverse clips with native audio up to 20 seconds in a single generation. The listed capabilities include text-to-video, image-to-video, video-to-video, audio-video continuation, keyframe transitions, multilingual dialogue, typography, broad style control, and chaining clips into longer multi-shot sequences.
The evaluation section is careful but ambitious. BFL says the current model and harness are still in development, and that the reported results are preliminary. In early 10-second 720p text-to-video comparisons, the company reports preference wins over several named video models, while also saying it expects further improvement during early access.
The more strategic detail is action. BFL says FLUX 3's world understanding extends to action prediction through two routes: native action prediction in the model itself, and specialized action models fine-tuned from the pretrained video backbone. Its FLUX-mimic partnership with mimic robotics is positioned around dexterous manipulation and production deployment.
The community angle is less about whether a demo clip looks impressive and more about whether generation, perception, and action can live on the same foundation without becoming a vague marketing umbrella. The release plan also matters: video and audio generation, action prediction, image synthesis, and an open-weight multimodal backbone are planned through early access phases.
Related Articles
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.
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.
AdaJEPA targets a practical weakness in robot and agent planning: frozen world models drift under distribution shift. The method updates during MPC with one gradient step per replan and a buffer of five recent transitions.