DOVEOptions
Options & Configuration · AiDotNet.Video.Options
Configuration options for DOVE, efficient one-step diffusion for real-world video super-resolution.
Inheritance: ModelOptions → NeuralNetworkOptions → DOVEOptions
For Beginners
DOVE upscales video by asking a large video generation model to clean up a blurry enlargement of it in a single step, which makes it far faster than diffusion upscalers that take dozens of steps.
How It Works
Defaults follow Chen et al., NeurIPS 2025 (arXiv 2505.16239) and the release (zhengchen1999/DOVE), which fine-tunes
CogVideoX1.5-5B as diffusers ships it (THUDM/CogVideoX1.5-5B): the 42-layer expert transformer (48 heads of
64, 2x2x2 patches, a 512-wide time embedding, 4,096-wide T5 text embeddings of 226 tokens) and the causal 3D VAE
(128, 256, 256, 512 channels, 16 latent channels, x8 in space and x4 in time, latent scale 0.7). The x4 bilinear
upscale of the clip is encoded and denoised in one v-prediction step at timestep 399 of the zero-terminal-SNR
scaled-linear schedule, conditioned on the pre-encoded empty prompt. Training is AdamW (β 0.9, 0.95, weight decay 1e-4, gradients clipped
to 0.1, a 100-step linear warmup): 10,000
latent-space MSE iterations at 2e-5, then 500 pixel-space iterations at 5e-6 on MSE, DISTS and the frame-difference
loss (1, 1), 80% of them on single frames.
Constructors
| Constructor | Summary |
|---|---|
DOVEOptions() |
Initializes a new instance with default values. |
DOVEOptions(DOVEOptions 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 BetaEnd { get; set; } |
Gets or sets the scaled-linear schedule’s last β (0.012). |
double BetaStart { get; set; } |
Gets or sets the scaled-linear schedule’s first β (0.00085). |
double DistsWeight { get; set; } |
Gets or sets the refinement stage’s DISTS weight λ1 (1; 0 leaves the VGG16 out). |
double FrameDifferenceWeight { get; set; } |
Gets or sets the refinement stage’s frame-difference weight λ2 (1). |
double ImageRatio { get; set; } |
Gets or sets the share of refinement iterations trained on a single frame (0.8). |
double LatentScale { get; set; } |
Gets or sets the latent scale (0.7). |
double LearningRate { get; set; } |
Gets or sets the adaptation stage’s AdamW rate (2e-5). |
double MaxGradientNorm { get; set; } |
Gets or sets the gradient norm clip (0.1, the release’s training scripts). |
double NormEpsilon { get; set; } |
Gets or sets the transformer’s LayerNorm epsilon (1e-5). |
double RefinementLearningRate { get; set; } |
Gets or sets the refinement stage’s AdamW rate (5e-6). |
double WeightDecay { get; set; } |
Gets or sets AdamW’s decoupled weight decay (1e-4, the release’s). |
int AdaptationIterations { get; set; } |
Gets or sets the latent-space adaptation iterations before pixel-space refinement (10,000). |
int AttentionHeadDim { get; set; } |
Gets or sets each head’s width (64); a multiple of 16 so the rotary embedding splits into pairs. |
int LatentChannels { get; set; } |
Gets or sets the latent channels (16). |
int MaxTextSeqLength { get; set; } |
Gets or sets the text tokens (226). |
int NumAttentionHeads { get; set; } |
Gets or sets the transformer’s attention heads (48). |
int NumFrames { get; set; } |
Gets or sets the frames of a clip (25, the paper’s training length). |
int NumLayers { get; set; } |
Gets or sets the transformer blocks (42). |
int NumTrainTimesteps { get; set; } |
Gets or sets the schedule’s training timesteps (1,000). |
int PatchSize { get; set; } |
Gets or sets the spatial patch side (2). |
int PatchSizeT { get; set; } |
Gets or sets the temporal patch length (2, CogVideoX 1.5’s). |
int ScaleFactor { get; set; } |
Gets or sets the spatial upscaling factor (4). |
int TemporalCompressionRatio { get; set; } |
Gets or sets the VAE’s temporal compression (4, a power of two). |
int TextEmbedDim { get; set; } |
Gets or sets the text embeddings’ width (4,096, T5-XXL’s). |
int TimeEmbedDim { get; set; } |
Gets or sets the time embedding’s width (512). |
int Timestep { get; set; } |
Gets or sets the timestep the one denoising step runs at (399). |
int VaeLayersPerBlock { get; set; } |
Gets or sets the VAE encoder’s residual blocks per stage (3; the decoder has one more). |
int WarmupSteps { get; set; } |
Gets or sets the linear warmup steps of each stage’s otherwise constant rate (100, the release’s). |
int[] VaeChannels { get; set; } |
Gets or sets the causal VAE’s stage widths (128, 256, 256, 512); each a multiple of 32. |
string ModelPath { get; set; } |
Gets or sets the path to the ONNX model file. |