Bora V1

Smallest video generation model from the @Codeically series. Flow-matching video DiT (336M parameters, dim 1024, depth 20) over frozen Wan2.1-VAE latents. 256x256, 8 fps, blocks of 4 latent frames with 1 frame of overlap, 8 sampling steps. Trained 50,000 steps at batch 32 on one A100-80G, on Ami1ne/Bora (Pexels / Pixabay / NASA clips). Text encoder: google/t5-v1_1-base, 32 tokens.

ema_latest.safetensors: EMA weights (ema_step50000.safetensors), bf16
checkpoints/ckpt_last.pt: full training state to resume
code/: bora_dit.py (model + sampler), bora_train.py, bora_dataset.py, bora_run.py
samples/: sample videos every 2,000 steps; logs/: training and validation curves

Generate: python code/bora_run.py sample --runs <dir with ema file> --prompts "ocean waves crashing". Decode latents with the Wan2.1 VAE: vae.decode(z * latents_std + latents_mean).

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support