Runway details how it made Gen-4.5 real-time
Runway has published the post-training recipe it uses to turn Gen-4.5 into a streaming, frame-by-frame video generator, betting that instant output will reshape how video gets made and what it costs.
Runway described its research into real-time video generation in a post that frames time-to-first-frame as the key metric, not total clip length or render time. The post follows last week's releases of Solaris and GWM Worlds 2. It does not launch a new model, and it gives no benchmarks or latency figures.
Runway starts from its existing base models, such as Gen-4.5, and post-trains them. Each autoregressive step is conditioned on an initial frame and a caption the user picks at the start. All generated latents stay in context to preserve quality, and the video and audio decoders run causally, so output streams out as it is produced.
The conversion happens in two stages. In the first, "teacher forcing," Runway turns the full-attention architecture into a temporally causal, frame-by-frame generator. In the second, "student forcing," it uses distribution matching distillation to cut the number of denoising steps per frame to a few. A frozen bidirectional teacher and a critic model guide the student. Distillation starts off-policy, which keeps memory use and step time low. According to Runway, most of the gains come from the on-policy phase that follows, where the student trains on its own rollouts and learns to correct drift instead of compounding it.
Runway also reports that a curriculum that gradually lengthens training sequences works better than a fixed sequence length. Starting with long rollouts let student trajectories drift so far from the teacher that its supervision signal became unstable.
The company says real-time generation moves the compute bottleneck from training to inference. Its post breaks off while describing the serving stack, so details there are limited.