Other Advanced Settings
Progress bar parameter instructions
During the task execution process, the progress parameters in the task output panel are as follows:
| Abbreviation | Explanation |
|---|---|
| R | The number of frames currently rendered |
| C | The number of frames currently being frame-filled or processed |
| RPT | ReadProcessTime single-frame read processing time |
| WPT | WriteProcessTime single-frame compression processing time |
| S | The most recently recognized transition |
| SC | The number of recognized transitions |
| TAT | TaskAcquireTime single frame-filling task acquisition time |
| PT | ProcessTime single frame-filling task processing time |
| QL | QueueLength frame-filling task queue length |
| SR | The number of frames currently being super-resolved |
| SRTAT | SuperResolutionTaskAquireTime single super-resolution task acquisition time |
| SRPT | SuperResolutionProcessTime single super-resolution task processing time |
| SRL | SuperResolutionQueueLength super-resolution task queue length |
Special instructions for progress bar parameters
In version 6.x and above, the "Decoding" and "Compression Pressure" panels in the output status panel will show the current decoding and compression pressure. When the decoding pressure and compression pressure are both 0, it means that there is currently no performance bottleneck and it is in the best state. The greater the values of the decoding pressure and compression pressure, the more serious the current bottleneck is. Note This value is inaccurate when using multiple frame-filling or super-resolution threads.
When only frame-filling without super-resolution, if the QL value remains below 10 for a long time, it indicates that the frame-splitting encounters a bottleneck. At this time, TAT>0.1. Please check whether the CPU usage is 100%. If so, please change the CPU software encoding compression parameters (such as changing the compression preset from slow to fast) or replace the CPU with更强 single-core performance. If QuickExtract is not enabled and the RPT value is large (>10s), please enable QuickExtract to reduce the CPU load. When enabling the high-precision workflow and CPU compression, if the CPU is not strong enough, QL=0 is a normal phenomenon.
When both super-resolution and frame-filling are performed, QL=0 and TAT>1 are normal phenomena. Try to ensure that SRL>0 and SRTAT to maximize the working efficiency of the graphics card, because SVFI performs super-resolution first and then frame-filling, for the same reason as above.
If the TAT (SRTAT) value is 0 and the QL (SRL) value is stable at a relatively high value, it means that the bottleneck of the working speed is mainly in the performance of the graphics card. This state can give full play to the performance of the graphics card, which is an ideal working situation.
Additional global settings
Note The current global settings correspond to version SVFI 6.15.9This type of global settings is set in the global_advanced_settings.json file (referred to as GAS) in the software root directory and can be edited by opening it with a text editor. By default, this file needs to be created manually.
The following is an example of all configurations. The value of the key-value pair is the default value. You can directly copy the following content and edit it as your own GAS file.
When actually editing the json file, the value corresponding to the non-existent key uses the default value:
{
"is_cudnn_enabled": true,
"trt_max_workspace_size": 1073741824,
"calib_sr_len": 750,
"is_cuda_graph_enabled": false,
"is_torchscript_enabled": false,
"is_cuda_loading_lazy": true,
"is_torch_trt_io32": false,
"is_rife_trtexec_prec_obey": false,
"show_trt_debug": false,
"debug_scdet": false,
"fmnet_model": "final",
"deep_deband_model": "deepdeband-f.pth",
"sisr_state_dict_key": "params",
"pure_scene_len_threshold": 20,
"use_minterpolate": false,
"vfi_use_pad": false,
"forw_dedup_use_pad": false,
"forw_dedup_use_correction": false,
"use_full_vspipe_seg": true,
"is_softsplat_stable": false,
"is_softsplat_gaussian": false,
"is_evict_flicker": false,
"use_xformers": true,
"drba_swap_thres": 1.0,
"torch_inference_mode": false,
"cv2_inter": "cubic",
"encoded_audio_format": "aac",
"encoded_audio_bitrate": "640k",
"custom_encoding": "utf-8",
"gui_scan_interval": 10
}
| Key Name | Explanation |
|---|---|
| is_cudnn_enabled | Whether to enable cudnn acceleration. |
| trt_max_workspace_size | The maximum workspace size of TensorRT in bits. When using vspipe for upscaling, it is not effective and you need to manually change the workspace size of vspipe.py. |
| calib_sr_len | The calibration times of TensorRT when using int8 quantization. |
| is_cuda_graph_enabled | Whether to enable CUDA Graph acceleration for supported models. |
| is_torchscript_enabled | Whether to enable TorchScript mode for supported models. |
| is_cuda_loading_lazy | Whether TensorRT enables lazy loading mode, which can reduce video memory usage. |
| is_torch_trt_io32 | Whether torch trt uses fp32 as the trt input and output. When enabled, it can significantly reduce the jagged edges caused by rife trt after forward deduplication. |
| is_rife_trtexec_prec_obey | Whether to use the precision constraints of RIFE trtexec. If enabled, add partial layer parameter range constraints to avoid possible image problems. |
| show_trt_debug | Whether to display TensorRT debug information. |
| debug_scdet | Whether to display debug information for scene detection. |
| fmnet_model | The FMNet model used, located in models/restorer/FMNet/models. |
| deep_deband_model | The DeepDeband model used, located in models/restorer/DeepDeband/models. |
| sisr_state_dict_key | The weight key of the custom SISR model used, applicable to ATD, SAFMN, SPAN, and Compact model structures. |
| use_minterpolate | Whether to use the minterpolate filter when reading videos with ffmpeg. Sometimes it can be enabled to correct the number of end frames. |
| pure_scene_len_threshold | Pure scene length threshold. |
| vfi_use_pad | Whether to use pad instead of interpolate for input resolution preprocessing in the frame interpolation process. |
| forw_dedup_use_pad | Whether to use pad instead of interpolate for input resolution preprocessing in forward deduplication. |
| forw_dedup_use_correction | Whether to use correction in forward deduplication. |
| use_full_vspipe_seg | When using the full vspipe workflow, automatically split the output chunks to support task progress replies. |
| is_softsplat_stable | Whether to enable SoftSplat stability support. |
| is_softsplat_gaussian | Whether to enable SoftSplat Gaussian kernel support. |
| is_evict_flicker | Whether to enable flicker removal. |
| use_xformers | By default, use xformer to accelerate the inference of certain models. |
| drba_swap_thres | DRBA swap threshold, with a value in (0,1). |
| torch_inference_mode | Whether to use torch.inference_mode() to accelerate inference. |
| cv2_inter | The cv2 interpolation method used. Optional values: nearest, linear, cubic, area, lanczos. |
| encoded_audio_format | Output audio format. |
| encoded_audio_bitrate | Output audio bitrate. |
| custom_encoding | The custom encoding format displayed in the output log, such as utf-8, gbk, etc. |
| gui_scan_interval | Monitoring folder scanning interval in seconds. |
Shortcuts
Main page
| Shortcut | Explanation |
|---|---|
| Ctrl+1 | Open the settings panel |
| Ctrl+2 | Open the preview panel |
| Ctrl+3 | Open the output status panel |
| Ctrl+4 | Open the Steam settings panel |
| Ctrl+5 | Open the preference settings panel |
| Backspace | Go back one page |
| Ctrl+B | Open the basic settings panel |
| Ctrl+A | Open the advanced settings panel |
| Ctrl+P | Show/Hide advanced settings panel |
| Ctrl+S | Save the current settings |
| Ctrl+Alt+L or Ctrl+Shift+L | Load default settings |
| Ctrl+D | Duplicate the current project |
| Delete | Delete the current project |
| Ctrl+Delete | Clear the project list |
| Ctrl+I | Import project |
| Ctrl+Return | Start frame-filling |
| Ctrl+E | Start compression |
- Holding down Ctrl and dragging in a folder will cause SVFI to automatically import all video files under that folder.
- Holding down Shift and dragging in a folder will cause SVFI to import that folder as a monitored folder.
Task status page
| Shortcut | Explanation |
|---|---|
| Ctrl+End | Stop frame-filling/compression |
| Ctrl+Space | Pause frame-filling/compression |
Mask drawing page
| Shortcut | Explanation |
|---|---|
| Ctrl+Z | Undo |
| Ctrl+R | Redo |
Player component
| Shortcut | Explanation |
|---|---|
| Space | Play/Pause |
| Right | Fast forward |
| Left | Fast backward |