SeedVROptions

Options & Configuration · AiDotNet.Video.Options

Options for SeedVR: the authors’ released SeedVR-7B (ByteDance-Seed/SeedVR, configs_7b/main.yaml), a window-attention MM-DiT on a causal video VAE.

Inheritance: ModelOptions → NeuralNetworkOptions → SeedVROptions

For Beginners

the defaults build the released 7-billion-parameter model. Shrink the widths and depth for experiments on modest hardware.

How It Works

The defaults are the release’s configuration: the 36-block NaDiT at width 3072 with 24 heads of 128, windows counted at 720p as (4, 3, 3) alternating with their half-shifted copies, the 16-channel VAE with 8x spatial and 4x temporal compression, the precomputed positive and negative prompt embeddings (58 and 64 tokens of 5120), classifier-free guidance and 50 Euler steps of the rectified flow. The paper’s own experiments use a 2.48B model of the same design initialized from SD3-Medium; set ModelDim, NumHeads, HeadDim and NumLayers for it.

Constructors

Constructor Summary
SeedVROptions() Creates the released configuration.
SeedVROptions(SeedVROptions other) Copies other.

Properties

Property Summary
OnnxModelOptions OnnxOptions { get; set; } The ONNX runtime options.
bool QkBias { get; set; } Whether the q, k, v projections have biases (released: false).
double ConditionNoiseScale { get; set; } The LR condition’s noise augmentation at inference, as a fraction of the schedule (the release’s script: 0.1).
double GuidanceRescale { get; set; } The guidance rescale (released: 0).
double GuidanceScale { get; set; } The classifier-free guidance scale (the release’s inference script: 6.5); 1 runs the positive prompt only.
double LatentScaling { get; set; } The latent scaling factor (released: 0.9152).
double LearningRate { get; set; } AdamW’s learning rate (SD3’s 1e-4; the paper follows SD3’s training settings).
double NormEpsilon { get; set; } The RMS norms’ ε (released: 1e-5).
double TrainingConditionNoiseScale { get; set; } The LR condition’s noise augmentation in training (released config: 0.25).
double WeightDecay { get; set; } AdamW’s weight decay (SD3 states none; 0).
int ExpandRatio { get; set; } The MLP’s expansion (released: 4).
int HeadDim { get; set; } The head width (released: 128).
int LatentChannels { get; set; } The latent channels (released: 16).
int ModelDim { get; set; } The DiT width, vid_dim = txt_dim (released 7B: 3072).
int NegativeTextLength { get; set; } The negative prompt’s tokens (the released neg_emb.pt: 64).
int NumFrames { get; set; } The frames of a clip.
int NumHeads { get; set; } The attention heads (released 7B: 24).
int NumLayers { get; set; } The blocks (released 7B: 36).
int SamplingSteps { get; set; } The Euler steps over the uniform trailing timesteps (released: 50).
int ScaleFactor { get; set; } The upscaling factor (4).
int TextDim { get; set; } The prompt embeddings’ width (released: 5120).
int TextLength { get; set; } The positive prompt’s tokens (the released pos_emb.pt: 58).
int VaeGroups { get; set; } The VAE’s group-norm groups (released: 32).
int VaeLayersPerBlock { get; set; } The encoder’s residual blocks per stage; the decoder has one more (released: 2).
int WindowColumns { get; set; } The windows across the width at 720p (released: 3).
int WindowFrames { get; set; } The windows along time at 720p (released: 4).
int WindowRows { get; set; } The windows down the height at 720p (released: 3).
int[] VaeChannels { get; set; } The VAE’s four stage widths (released: 128, 256, 512, 512).
string ModelPath { get; set; } The ONNX model’s path in inference mode.