🎬 VSpark — 0.5B Video Generation Model

VSpark is a fully trainable, scratch-initialized 726.2M parameter video generation model built with a Spatial-Temporal Diffusion Transformer architecture.

Architecture

Component	Details
Video backbone	Spatial-Temporal DiT · 16 blocks · dim=1024 · 16 heads
Video VAE	4-level encoder/decoder · latent_ch=4 · stride-8
Text encoder	6-layer transformer · dim=768 · BPE vocab=49408
Audio decoder	8-block mel-spectrogram DiT · dim=512
Total	~726.2M params

Quick start

from vspark_pipeline import VSparkPipeline
pipe = VSparkPipeline.from_pretrained("AyaanAhmed123/Vspark")
frames, audio_mel = pipe("A golden sunset over ocean waves", num_inference_steps=50)

Training

python train_colab.py --data_dir /path/to/videos --epochs 50

License

Apache 2.0