One model for pixels, frames, and sound — the generative stack collapses into a single object

The image labs are adding video, the language labs are adding generation, and the short-video companies are unifying audio and picture. Every path is converging on the same destination.

Black Forest Labs announced FLUX 3, its first multimodal frontier model, and Meta launched Muse Image while previewing Muse Video. The image labs, which built their names on one modality, are now declaring frontier multimodal systems. The boundary between an image model and a full multimodal system is dissolving from the image side.

Convergence from every direction

It is dissolving from the other side too. ByteDance's Seedance 2.0 generates synchronised audio and video from one unified architecture, and Alibaba's Qwen previewed a trillion-parameter multimodal model processing text, images, video, and documents. A short-video company unifying sound and picture, a cloud lab scaling an omni-model past a trillion parameters, image labs claiming frontier video — all in one window, all aimed at the same object.

The object they are converging on

That object is a model that takes prompts across modalities and outputs pixels, frames, sound, or all three. For years image models, video models, and language models were separate research programmes with separate architectures. The 2026 turn is the recognition that they are facets of one capability — predicting how a scene looks, moves, and sounds — and can share a model.

The unified-audio-video piece is the technically hardest and most telling. Generating synchronised sound and picture from a single model, rather than dubbing audio onto generated video, means the model has learned the joint distribution of what a scene looks and sounds like. That is a genuinely richer object than either modality alone, and it is the direction the whole field is moving.

Where this points

The same convergence shows up on the research side as action models and world models — systems where generating a video of a motion and executing that motion become two readouts of one model. As the generative stack collapses into a single multimodal object, the line between a model that depicts the world and one that acts in it is the next boundary to go. The image-video merger this cycle is a step on that longer path.

Digital Applied — Seven days, seven model releases: the new AI normal → · ThursdAI — July 2026 AI releases: OpenAI, Anthropic, Google DeepMind, Cognition →