The Runway video starts playing before it is fully generated

Runway details how its models are becoming causal and autoregressive in order to display the first frames while the rest of the video is still being generated.

A video starts playing before the model has finished generating it. Instead of entering a prompt, waiting for the entire clip to be calculated, and then viewing the result, Runway wants to display the first frames immediately while producing the rest as a continuous stream.

In a research post published on September 10, 2026, the company outlines the work behind this approach. It is not announcing a new commercial model or a feature that users can activate today. The post describes a broader research direction already reflected in Runway Characters, Solaris, and GWM Worlds 2.

The main target is time to first frame. Most current video generators treat each output as a finished object. The user submits a prompt, waits several seconds or minutes, and receives a complete sequence. If the result does not match the original intent, the process begins again with a new generation.

Runway is trying to replace that sequence of waiting and restarting with a continuously observable process. The first images would appear while the model is still deciding what comes next. In principle, shorter feedback loops would let a creator spend more time directing a scene and less time waiting for successive versions.

The word “instant” should be treated as a goal rather than a measured specification. Runway does not provide an exact time to first frame, playback rate, resolution, hardware configuration, public release date, API, or price for the prototype described in the post. The demonstrations show streaming generation, but they do not establish how quickly or economically the same system could serve many customers at once.

The method begins with two inputs: an initial frame and a caption selected at the start of generation. The model then produces the video frame by frame, using its previous outputs as context for the next ones. Generated latent representations remain available to help maintain visual continuity.

The video and audio decoders also operate causally. They do not need to wait for the complete sequence of latent states before beginning their work. As new states become available, the decoders can turn them into audiovisual output and stream the result.

This differs from the full-attention approach used by many high-quality video models. Those systems can examine an entire sequence at once and coordinate distant moments before producing the final clip. A real-time model cannot depend on information from frames that do not yet exist. It must commit to each stage while preserving enough context to continue coherently.

Runway says it is adapting existing foundation models such as Gen-4.5, rather than training a separate architecture from the ground up for every real-time application. The company first converts the base model into a temporally causal, autoregressive generator.

It calls this stage “teacher forcing.” During training, the model learns to predict the next state from correct earlier states supplied by the dataset. Runway says the architecture retains the strengths of the original bidirectional model after this conversion, although the post provides no benchmark comparing the causal version directly with Gen-4.5.

Making generation causal does not make it fast enough to stream. The underlying flow-matching process still requires many denoising steps to produce each frame. Runway therefore distills the larger model into a student that can generate each frame in only a few steps.

The company uses distribution-matching distillation for this compression. A frozen bidirectional teacher indicates what high-quality output should look like, while a critic tracks the distribution produced by the student. Training pushes the student’s predictions closer to those of the teacher without requiring it to reproduce the teacher’s complete calculation process.

The first distillation phase is described as off-policy. The student predicts the next state from ground-truth context rather than from its own previous generations. This arrangement reduces training memory requirements and keeps each optimization step relatively short.

It also creates a gap between training and actual use. During inference, the model does not receive a perfect history. Every new frame depends on images it generated itself. A minor error can enter the context, influence the next prediction, and grow over time.

This problem is particularly visible in video. A language model can sometimes correct a mistake in a later sentence. A malformed shape, altered face, or misplaced object in one image becomes part of the visual evidence used to produce the next. Small defects can gradually turn into large deformations.

Off-policy training does not expose the model to enough of these self-produced errors. Runway therefore adds an on-policy phase in which the student performs complete autoregressive rollouts during training. Each generated latent state becomes the context for the next step, matching the conditions it will encounter when deployed.

The model can then learn how its own errors develop and how to limit their amplification. Runway says this on-policy stage delivers most of the practical gains in the real-time generation pipeline, although it does not quantify that contribution separately.

Training long rollouts immediately proved unstable. Errors accumulated until the student’s trajectory diverged so far from the teacher that the teacher’s target no longer provided useful guidance. The expected output and the sequence actually being generated had become disconnected.

Runway addressed the problem with a curriculum that gradually increases sequence length. The student begins with shorter trajectories, then learns to sustain longer ones as training progresses. According to the company, this approach performs better than using one fixed sequence length throughout training.

The post does not fully explain how the system manages context during an extended session. Keeping every generated latent state available can preserve quality, but it also causes memory and