diff --git a/docs/sana_streaming.md b/docs/sana_streaming.md index 3b4e9b4a..b58f9c21 100644 --- a/docs/sana_streaming.md +++ b/docs/sana_streaming.md @@ -219,8 +219,12 @@ python inference_video_scripts/v2v/inference_sana_streaming.py \ The release includes three source videos under [`Efficient-Large-Model/SANA-Streaming`](https://huggingface.co/Efficient-Large-Model/SANA-Streaming/tree/main/source). -The same examples can be run with both `long_streaming` and -`bidirectional_short` by changing `--mode`. +The same source videos and prompts can be used with both `long_streaming` +and `bidirectional_short`. If you rely on mode-specific defaults, omit +`--config`, `--model_path`, `--num_frames`, `--step`, and `--cfg_scale`: the +script selects those defaults from `--mode`. When copying either full command +above, update all of these explicit arguments together; changing only `--mode` +does not replace an explicitly supplied config, checkpoint or sampling setting. | Example | Source video | Prompt | |---------|--------------|--------| @@ -241,7 +245,7 @@ The same examples can be run with both `long_streaming` and | `--model_path` | DiT checkpoint path. Defaults to the released Hugging Face checkpoints. | | `--num_frames` | Frames decoded from the source video (`969` for streaming, `81` for bidirectional). | | `--height / --width` | Center-cropped output resolution (`704 × 1280`). | -| `--fps` | Output MP4 frame rate (`16`). | +| `--fps` | Saved MP4 playback frame rate (`16`); does not set or measure inference throughput. | | `--step` | Denoising steps (`4` for streaming, `50` for bidirectional). | | `--cfg_scale` | CFG scale (`1.0` for streaming, `6.0` for bidirectional). | | `--flow_shift` | Optional scheduler flow-shift override. | @@ -250,6 +254,15 @@ The same examples can be run with both `long_streaming` and | `--num_cached_blocks` | Streaming cache window size (`2`). | | `--sink_token` | Keep the first chunk in the streaming cache window (`true`). | +### Interpreting playback FPS + +The script reads frames from `--video_path`, runs inference, then passes +`--fps` to the output video writer. Raising this value makes the saved frames +play faster; it does not make the model generate more frames per second. For a +performance report, record the generated frame count, elapsed inference time, +hardware and sampling configuration separately. The paper's deployment figures +above and the saved file's playback clock measure different things. + ## 📁 HF Repository Layout `Efficient-Large-Model/SANA-Streaming_bidirectional`: