Video glossary
The words in our tests, in plain language. Every term has its own link so a review can point straight to the definition.
t2v · text-to-video
Generate a video from a written description. The model supplies the images and motion; a prompt alone does not guarantee the exact action, timing, or text you intended.
i2v · image-to-video
Generate motion using an input image as visual conditioning, often as a first frame. It helps establish the scene or character, but identity can still drift during the clip.
Image + audio
Condition a video on both a reference image and an audio track. The picture guides appearance and the sound can guide speaking motion. Accepting audio is not proof that the lips match it.
Lip-sync
How well visible speech movements match the timing and sounds of the spoken audio. A moving mouth is not enough: check speech starts, pauses, consonants, and the end of a line.
CFG · classifier-free guidance
A way to steer generation by combining predictions made with and without conditioning, such as a text prompt. Guidance scale controls the strength of that steering. Higher values are not always better, and the useful range depends on the model and workflow.
Seed
A number used to initialize a random-number generator. Keeping it fixed helps compare runs when the model, settings, and software also stay fixed. The same seed does not guarantee identical output across hardware or software versions.
Distilled
Trained to reproduce useful behavior from a teacher model or a longer generation process. Video diffusion distillation often aims to use fewer sampling steps. Follow the distilled checkpoint's own settings; fewer steps do not make every result equally good.
LoRA · low-rank adaptation
A compact set of trained changes applied to a compatible base model. A LoRA can teach a style, subject, task, or faster sampling behavior. It is usually not a complete standalone model, and its strength and base-model compatibility matter.
Frame interpolation
Create intermediate frames between existing frames. It can make motion look smoother or support slow motion, but it does not recover the true missing motion and can create warped edges or double images.
Upscaling
Increase an image or video's pixel dimensions. A learned upscaler may synthesize plausible detail; those details are not evidence of what was originally present. Check faces, text, and consistency from frame to frame.
Staged weights
In this lab, model-weight files already downloaded and placed where the workflow expects them. Staged does not mean loaded into RAM or GPU memory, and it does not prove the workflow has run successfully.
MemAvailable
Linux's estimate of how much memory is available for starting new applications without swapping. It includes some reclaimable memory, so it differs from MemFree. A sampled minimum can miss a brief dip; it is not a measurement of a model's peak allocation.
Unified memory
On the DGX Spark, the CPU and GPU share physical system memory. Model weights, video buffers, desktop apps, and the operating system compete for the same pool. Installed capacity is not the amount available to a job.
Sampling steps
The iterations used to turn noise into generated output. More steps take more compute and may change the result, but do not guarantee higher quality. A distilled model may be trained for a small, specific step count.
Quantization
Store or compute model values at reduced numerical precision to save memory or improve speed. Support and visual impact depend on the model and runtime. A smaller file does not by itself establish the memory needed for a full job.
Cold and warm runs
A cold run may include loading a model or compiling code. A warm run reuses some preparation. Report what the timer includes; a first-run delay is not the same as steady-state generation speed.
Seconds per video second
Elapsed compute or wall time divided by the output video's duration. At 15 s/s, a five-second clip takes about 75 seconds within that timer's scope. State whether loading, polling, or post-processing is included.
Temporal consistency
How well appearance and structure hold together over time: a face remains the same face, clothing keeps its pattern, and objects do not flicker or transform between frames.