Glossary of realtime video and world model terms
Definitions for realtime video and world models: sessions, chunks, autoregressive diffusion, KV cache, visual drift, transport and ready-second billing.
The vocabulary of realtime video generation, in the order a developer tends to meet it. Definitions are short on purpose; the guides go deeper.
Terms
Realtime model
A model that produces output continuously while it is being steered, at a pace a viewer can watch live. It differs from a batch model, which takes a request and returns a finished file after a wait. Realtime video models, interactive world models and live video editors are all realtime models.
Realtime video generation
Generating video fast enough, and in small enough pieces, that frames can be delivered to a viewer while later frames are still being computed. It usually relies on autoregressive generation in chunks and on a client that buffers a short playout window.
Interactive video
Video that responds to input while it plays: a changed prompt, a camera move, a key press or a webcam frame. The response arrives within seconds or less, not after a re-render.
World model
A generative model that simulates an environment frame by frame, conditioned on what happened before and on the actions taken. A world model produces a scene you can act in rather than a clip you only watch.
Navigable world model
A world model controlled by movement and look controls, typically WASD keys plus a camera direction, usually starting from a seed image. Examples on Windflow are LingBot-World, Matrix-Game, DreamX-World and ABot-World.
Seed image
The starting image a world model or image-to-video model is conditioned on. The first frames, the style and the setting of the generated world follow from it.
Session
A held GPU assigned to one running model. You connect once, send controls and receive frames over the same connection while the model keeps running, then release it. A session is the unit that is started, steered and billed.
Ready session-second
Windflow's billing unit: one second during which a session is ready and returning frames. Queueing and worker startup are not counted. Rates are published per model; see pricing.
Track
A named stream a session sends or receives. Windflow models return video on main_video and, for models that generate sound, audio on main_audio. Webcam input travels on a send-only track.
Prompt steering
Changing a text prompt while a session runs so the generated video follows the new direction. Depending on the model, the change blends in softly or waits for the next chunk or window boundary.
Soft transition and scene cut
A soft transition keeps the model's rolling history, so subjects, layout and camera motion continue as the prompt changes. A scene cut discards that history to start independent content, which is cleaner for a new shot but costs a restart of the history.
Autoregressive video generation
Generating video in sequence, each chunk conditioned on what was already produced, instead of generating every frame at once. It is what lets a stream continue indefinitely and accept new input between chunks.
Chunk, block and window
The unit a model generates at a time: a fixed number of frames, such as 12, 24 or 33, or a longer window of several seconds. Controls sent during generation typically take effect at the next boundary, so chunk length sets how quickly a model reacts.
Diffusion transformer (DiT)
A diffusion model whose denoising network is a transformer operating on video latents. Most current video generators, including Helios, LongLive, LingBot-World and LTX, are built this way.
Denoising step
One pass of the denoising network. Standard diffusion needs many; realtime models are distilled to a handful, which is the main lever on how fast a chunk is produced.
Distillation
Training a model to reach in a few steps what a teacher model reaches in many. Techniques named in realtime video work include Self-Forcing, which trains the model on its own autoregressive output, and DMD2, a few-step distribution-matching distillation.
KV cache and rolling history
The stored attention keys and values of frames already generated. An autoregressive model reuses them so each chunk does not recompute the past; a rolling cache drops the oldest frames to keep memory bounded, which is why very old content can be forgotten.
Drift
Gradual degradation over a long autoregressive run: identity, colour or geometry slowly wander because each chunk inherits the previous chunk's small errors. Long-running models counter it with anchoring, memory and re-conditioning.
Re-anchoring
Periodically refreshing a long run with a clean conditioning signal, such as the original reference frame, to pull it back toward the intended look. A refresh can be visible as a brief jump.
Time to first frame
The delay between starting a session and the first frame reaching the viewer. For windowed models it can include a whole window of generation, so it is longer than the steady-state delay between frames.
Control latency
The delay between sending a control and seeing its effect on screen. It has two parts: how long until the next chunk boundary, and how long the stream takes to reach the viewer.
Playout buffer
A short queue on the client that smooths uneven frame arrival. A bigger buffer hides stalls and adds latency; a smaller one does the opposite.
WebRTC
A real-time media protocol built on UDP, with congestion control, loss recovery and browser support. It suits clean links and interactive audio and video; on a lossy path, keyframe recovery can cause visible freezes.
WebSocket (H.264)
A single TCP connection carrying encoded video frames and control messages. It is simple to traverse and to debug, and it trades WebRTC's loss handling for in-order delivery.
Super-resolution
A learned upscaler that turns frames generated at a lower resolution into the delivery resolution, for example 896 × 512 into 1344 × 768. It lets a model generate faster and still deliver more detail.
Image-to-video and text-to-video
Image-to-video starts from a reference image; text-to-video starts from a prompt alone. Many realtime models accept either, and some take only one.
Joint audio-video generation
Generating sound together with the picture in one model, so dialogue, ambience and music follow the scene, as opposed to dubbing audio onto finished video afterwards. LTX-2.5 and MiniMax FastH3 work this way.
Open weights
A model whose trained parameters can be downloaded and run by anyone, under the licence its authors chose. Open weights are not the same as an unrestricted licence: terms on commercial use, territory and disclosure vary by model.