DualXVSROptions
Options & Configuration · AiDotNet.Video.Options
Configuration options for DualX-VSR, the dual axial spatial x temporal transformer for real-world video super-resolution without motion compensation.
Inheritance: ModelOptions → NeuralNetworkOptions → DualXVSROptions
For Beginners
DualX-VSR upscales video four times without estimating motion. These settings control how wide and deep its attention is and how it is trained: first to match clean frames, then also to look realistic.
How It Works
Defaults follow Cao et al. 2025 (arXiv 2506.04830), Tables 6-8 and Sec. 5.1: a 64-channel 3x3 shallow convolution,
2x2 patches embedded to 1,280 channels, six vertical-temporal then six horizontal-temporal blocks, 12 heads, MLP ratio
2, RoPE with base 10,000, temporal transformers of depth 2 in the embedding and the reconstruction, x4 upsampling and
16-frame clips. Training is AdamW at 5e-5 for 500,000 bicubic pretraining and 150,000 real-degradation fine-tuning
iterations on Charbonnier loss, then 25,000 iterations at 1e-5 for the generator and discriminator on Charbonnier,
perceptual and GAN losses weighted 0.01, 1 and 0.005. The paper gives no code; HeadDim (64, vit-pytorch’s
ViViT default) is the head width that reproduces its 127.95M parameters, and AdamW’s unstated decay is PyTorch’s 0.01.
Constructors
| Constructor | Summary |
|---|---|
DualXVSROptions() |
Initializes a new instance with default values. |
DualXVSROptions(DualXVSROptions other) |
Initializes a new instance by copying from another instance. |
Properties
| Property | Summary |
|---|---|
OnnxModelOptions OnnxOptions { get; set; } |
Gets or sets the ONNX runtime options. |
double AdversarialLearningRate { get; set; } |
Gets or sets the generator’s AdamW rate in the GAN stage (1e-5). |
double AdversarialPixelWeight { get; set; } |
Gets or sets the Charbonnier weight in the GAN stage (0.01). |
double CharbonnierEpsilon { get; set; } |
Gets or sets the Charbonnier ε (1e-6, i.e. |
double DiscriminatorLearningRate { get; set; } |
Gets or sets the discriminator’s AdamW rate (1e-5). |
double GANWeight { get; set; } |
Gets or sets the GAN loss weight in the GAN stage (0.005). |
double LearningRate { get; set; } |
Gets or sets the AdamW rate of the pretraining and fine-tuning stages (5e-5). |
double MlpRatio { get; set; } |
Gets or sets the MLP’s hidden width as a multiple of the token width (2). |
double PerceptualWeight { get; set; } |
Gets or sets the perceptual loss weight in the GAN stage (1; 0 leaves the VGG19 out). |
double RopeBase { get; set; } |
Gets or sets the rotary position embedding’s base b (10,000). |
double WeightDecay { get; set; } |
Gets or sets AdamW’s decoupled weight decay (0.01, PyTorch’s default; the paper states none). |
int DiscriminatorChannels { get; set; } |
Gets or sets the discriminator’s base width (64, RealBasicVSR’s U-Net discriminator). |
int EmbedDim { get; set; } |
Gets or sets the dual axial transformer’s token width (1,280). |
int EmbeddingTemporalDepth { get; set; } |
Gets or sets the blocks of the input embedding’s temporal transformer (2). |
int HeadDim { get; set; } |
Gets or sets the channels of each attention head (64). |
int HorizontalDepth { get; set; } |
Gets or sets the horizontal-temporal attention blocks, which follow the vertical ones (6). |
int NumFeatures { get; set; } |
Gets or sets the shallow convolution’s channels and the per-pixel temporal transformers’ width (64). |
int NumFrames { get; set; } |
Gets or sets the frames of a clip (16). |
int NumHeads { get; set; } |
Gets or sets the attention heads of every transformer (12). |
int PatchSize { get; set; } |
Gets or sets the side of the square image patch a token covers (2). |
int ReconstructionIterations { get; set; } |
Gets or sets the Charbonnier-only iterations before the GAN stage (650,000: 500,000 pretraining plus 150,000 fine-tuning, which differ only in their data). |
int ReconstructionTemporalDepth { get; set; } |
Gets or sets the blocks of the reconstruction’s temporal transformer (2). |
int ScaleFactor { get; set; } |
Gets or sets the spatial upscaling factor, a power of two (4). |
int VerticalDepth { get; set; } |
Gets or sets the vertical-temporal attention blocks (6). |
string ModelPath { get; set; } |
Gets or sets the path to the ONNX model file. |