Docs · Questions
StreamDiffusionTDUsage
How do I enable StreamV2V / cached attention in StreamDiffusionTD v0.3.1?
StreamV2V / Cached Attention (v0.3.1)
StreamV2V adds temporal coherence to video-to-video generation using cached key-value attention. It was available in v0.2.99, missing in v0.3.0, and is back in v0.3.1.
Requirements
- TensorRT acceleration required (acceleration must be set to
tensorrt) - peft package (>= 0.6.0) must be installed. Run
python -m sd_installer verifyto check - Engine is resolution-locked: if you built your TRT engine at 512x512, cached attention only works at 512x512. Other resolutions need a new engine build
How to Enable
- Go to Settings 2 page on the operator
- Set Cattenable to On
- Adjust Cattmaxframes (number of frames in the attention cache)
- Adjust Cattinterval (how often the cache refreshes)
- Start the stream with TensorRT acceleration
VRAM
StreamV2V is relatively lightweight on VRAM. On a 4090 (24 GB), base SDXL-turbo with StreamV2V uses around 13 GB total (including Windows baseline).
Troubleshooting
- Error on enable: Make sure acceleration is set to
tensorrt, notnone - Visual artifacts at different resolution: Engine is locked to build resolution. Rebuild the engine at your target resolution
- Import error mentioning peft: Install the peft package in your venv:
pip install peft - Missing after upgrade from v0.3.0: StreamV2V was not in v0.3.0. If you are on v0.3.1 and don’t see the parameters, run
python -m sd_installer repairto ensure the diffusers fork is correct