RealisVSROptions

Options & Configuration · AiDotNet.Video.Options

Options for RealisVSR: Wan 2.1-1.3B with the Consistency-Preserved ControlNet, trained with the High-Frequency Rectified Diffusion Loss (Zhao et al. 2025).

Inheritance: ModelOptions → NeuralNetworkOptions → RealisVSROptions

For Beginners

the defaults build the paper’s 1.6B-parameter model. Shrink ModelDim, NumLayers, ControlNetLayers and the VAE widths for experiments on modest hardware.

How It Works

The paper names its backbone (Wan-1.3B, whose released configuration supplies the DiT and VAE defaults here), the loss weights, the HOG bins and the optimizer. It does not state the ControlNet’s depth, the HOG cell and block sizes, the sampler or the training timestep distribution, and the authors’ repository is no longer published; those defaults are documented on each property.

Constructors

Constructor Summary
RealisVSROptions() Creates the paper’s configuration.
RealisVSROptions(RealisVSROptions other) Copies other.

Properties

Property Summary
OnnxModelOptions OnnxOptions { get; set; } The ONNX runtime options.
double ConditionScale { get; set; } γ’s initial value, the paper’s condition scale (1.0).
double LearningRate { get; set; } AdamW’s learning rate (Sec.
double NormEpsilon { get; set; } The norms’ ε (1e-6).
double SigmaShift { get; set; } The flow-matching shift s, σ’ = s·σ / (1 + (s − 1)·σ).
double WeightDecay { get; set; } AdamW’s weight decay (Sec.
double[] WaveletWeights { get; set; } The Haar band weights {LL, LH, HL, HH} of the wavelet loss (Eq.
int ControlNetLayers { get; set; } L_CPC, the ControlNet branch’s blocks.
int FfnDim { get; set; } The feed-forward width (Wan-1.3B: 8960).
int FrequencyDim { get; set; } The sinusoidal time embedding’s width (Wan-1.3B: 256).
int HogBins { get; set; } The HOG loss’s orientation bins over [0°, 180°] (Eq.
int HogBlockSize { get; set; } The HOG block side in cells.
int HogCellSize { get; set; } The HOG cell side in latent pixels.
int ModelDim { get; set; } The DiT width (Wan-1.3B: 1536).
int NumFrames { get; set; } The frames of a training clip (49).
int NumHeads { get; set; } The attention heads (Wan-1.3B: 12).
int NumLayers { get; set; } L_main, the main DiT’s blocks (Wan-1.3B: 30).
int NumTrainTimesteps { get; set; } The training schedule’s length; training draws σ uniformly from its shifted values (1000).
int SamplingSteps { get; set; } The Euler steps of the rectified-flow sampler.
int ScaleFactor { get; set; } The upscaling factor (4).
int TextDim { get; set; } The umT5 embedding width the DiT reads (4096).
int TextLength { get; set; } The prompt’s token count (512).
int VaeDim { get; set; } The Wan VAE’s base width (96).
int VaeResidualBlocks { get; set; } The Wan VAE’s residual blocks per stage (2).
int[] VaeDimMultipliers { get; set; } The Wan VAE’s stage multipliers ({1, 2, 4, 4}: 8x in space, 4x in time).
string ModelPath { get; set; } The ONNX model’s path in inference mode.