Match the latency problem to the part of generation you can change

A visual generator may spend time refining an entire output, revisiting parts that already look settled, or waiting to finish a clip before showing anything. Those delays call for different changes.

Fewer denoising evaluations describe reduced computation; they do not by themselves establish a reduction in stored memory.

Across the tested systems, training shortened a shared decoding schedule, early acceptance reduced refinement for selected video tokens, and independently trained updates combined few-step generation with streaming.

The useful choice depends on the delay a product faces and on quality measured in that product’s setting.

Imagine an application that must describe an image promptly, then generate video as a conversation continues. A long sequence of decoding steps can delay the description.

In video, some regions may settle while others still change, and a clip that appears only after all its frames are complete cannot support the same interaction as a stream. The studies here examine these concrete sources of delay in different systems, so their results answer related design questions rather than forming one speed contest.

A decoding or denoising step is another pass in which a generator revises its proposed output. The text and image model studied here fills masked positions over successive steps.

The video model repeatedly refines estimates for portions of a new chunk. A third system produces audio and video chunks in sequence so a completed chunk can be emitted while later ones are still being made. The shared question is where changing that work helps under a particular latency constraint.

There are three places to intervene. Training can teach a model to cover its usual trajectory in fewer steps, changing the schedule used across an output.

During generation, a system can stop revising individual portions judged ready and keep working on the rest. Or a few-step model can gain chunkwise streaming so the first output arrives before the entire sequence is complete.

Each proposal changes a different part of generation, and each needs evidence at the task and budget where it will be used.