UniVSTOptions
Options & Configuration · AiDotNet.Diffusion.StyleTransfer
Configuration for UniVSTModel.
For Beginners
UniVST restyles only a chosen region of a video and leaves the rest alone. It does not train anything — it steers a pretrained image diffusion model. These settings control which region is tracked, how strongly the style is applied, and how much the result is smoothed across frames so it does not flicker.
How It Works
Defaults are the published values from Song, Lin, Zhan, Yan, Cao and Ji, “UniVST: A Unified Framework for Training-free Localized Video Style Transfer” (arXiv:2410.20084, TPAMI 2025): Stable Diffusion v1.5, T = 50 DDIM steps, 512x512, 16 frames per batch.
Every timestep boundary is a FRACTION of T rather than an absolute step index, because that is
how the paper states them and because it keeps the schedule meaningful when
DdimSteps changes. A literal step number would silently mean something different
at T = 20 than at T = 50.
Constructors
| Constructor | Summary |
|---|---|
UniVSTOptions() |
Properties
| Property | Summary |
|---|---|
double BetaAtRampEnd { get; set; } |
Gets or sets beta at the end of the ramp. |
double BetaAtRampStart { get; set; } |
Gets or sets beta at the start of the ramp. |
double KeyValueAdaInEndFraction { get; set; } |
Gets or sets the end of the key/value AdaIN ramp, as a fraction of T. |
double KeyValueAdaInStartFraction { get; set; } |
Gets or sets the start of the key/value AdaIN ramp, as a fraction of T. |
double LatentAdaInEndFraction { get; set; } |
Gets or sets the end of the latent-level AdaIN window, as a fraction of T. |
double LatentAdaInStartFraction { get; set; } |
Gets or sets the start of the latent-level AdaIN window, as a fraction of T. |
double MaskDownsampleRate { get; set; } |
Gets or sets the random downsampling rate applied to candidate anchor points. |
double MaskFeatureTimestepFraction { get; set; } |
Gets or sets the DDIM-inversion timestep, as a fraction of T, whose feature maps drive mask propagation. |
double QueryBlendGamma { get; set; } |
Gets or sets gamma, the query-blending weight between the stylized and content queries. |
double SmoothingEndFraction { get; set; } |
Gets or sets the end of the smoothing window, as a fraction of T. |
double SmoothingStartFraction { get; set; } |
Gets or sets the start of the smoothing window, as a fraction of T. |
int DdimSteps { get; set; } |
Gets or sets the number of DDIM steps (T). |
int MaskAnchorHistory { get; set; } |
Gets or sets how many preceding frames act as anchors alongside the first frame. |
int MaskFeatureUpBlockIndex { get; set; } |
Gets or sets which UNet upsampling block supplies the mask-propagation features. |
int MaskMatchNeighbors { get; set; } |
Gets or sets k, the number of nearest neighbours matched per point. |
int NumFrames { get; set; } |
Gets or sets the number of frames processed together. |
int SmoothingHalfWindow { get; set; } |
Gets or sets m, the half-width of the smoothing window; the full window is 2m + 1 frames. |
Methods
| Method | Summary |
|---|---|
void Validate() |
Validates the configuration and throws when a value cannot describe a usable schedule. |