Clip length sweeps
Each configuration is one model at one resolution and frame rate, run
at clip lengths of growing size. The prompt is encoded once and the text encoder
freed before the generator loads, so every clip is measured from denoise start to
file written. Produced by video.py.
Colours identify configurations, not models. The x axis is clip length in seconds of video.
Wall time per clip
Denoising plus VAE decode, prompt already encoded. Lower is better; log scale, because the fast tier and the 50-step models are two orders of magnitude apart. The curves bend upward because attention cost grows faster than the frame count.
Seconds of video per wall minute
The same clips as a rate. Above 60 is faster than real time.
Peak VRAM
Peak allocated during the clip, weights included. A clip that ran out of memory has no point here but keeps its row in the table.
Every clip
| Configuration | Frames | Clip | Wall | Per step | Decode | Video s / min | Peak VRAM | Status |
|---|
Configurations
What each sweep ran: the model, its shape, and how it was loaded.
| Configuration | Repository | Size | fps | Steps | Guidance | VAE tiling | Load | Encode | Longest fit |
|---|