NaviCache is a training-free, test-time self-calibration caching method for accelerating Open-Sora 1.2 text-to-video generation.
Clone NaviCache, install its dependencies, and download the Open-Sora 1.2 checkpoints:
git clone https://github.com/HelloZicky/NaviCache.git
cd NaviCache
pip install -r requirements.txt
huggingface-cli download hpcai-tech/OpenSora-STDiT-v3 \
--local-dir checkpoints/opensora/OpenSora-STDiT-v3
huggingface-cli download hpcai-tech/OpenSora-VAE-v1.2 \
--local-dir checkpoints/opensora/OpenSora-VAE-v1.2
huggingface-cli download DeepFloyd/t5-v1_1-xxl \
--local-dir checkpoints/opensora/t5-v1_1-xxlRun the paper configuration at 51 frames, 848x480, and 30 sampling steps:
python NaviCache4OpenSora/eval/navicache/experiments/opensora.py \
--transformer_path checkpoints/opensora/OpenSora-STDiT-v3 \
--vae_path checkpoints/opensora/OpenSora-VAE-v1.2 \
--text_encoder_path checkpoints/opensora/t5-v1_1-xxl \
--prompt_path NaviCache4OpenSora/eval/navicache/vbench/VBench_full_info.json \
--navicache_thresh 0.35 \
--navicache_align_steps 5The integration defaults select Open-Sora 1.2, 51 frames, 480p at a 9:16 aspect ratio (848x480), and 30 sampling steps.
| Method | Latency | Speedup | Hardware |
|---|---|---|---|
| Open-Sora 1.2 | 56.48 s | 1.00x | RTX 4090 |
| TeaCache | 41.38 s | 1.36x | RTX 4090 |
NaviCache (threshold=0.35, align_steps=5) |
35.29 s | 1.60x | RTX 4090 |
Prompt: a train accelerating to gain speed
Analysis: TeaCache turns the tall green train into a low blue-and-red streak with a mismatched car body, whereas NaviCache retains the Native train, track-side composition, and left-to-right trajectory.
We would like to thank the contributors to Open-Sora, VideoSys, and TeaCache.