Compared to the stable release, the current open beta includes:
Real-time super-resolution and frame generation for various windows (players, web pages, games)
Turbo Mode support for non-NVIDIA GPUs, based on DirectML backend, compatible with .onnx super-resolution models under the TensorRT category, as well as RIFE models
Instructions for real-time super-resolution and frame generation:
{STEAM_CLAN_IMAGE}/40952458/a0f5797d6218d7670a5dcd1c55d46cd2b3d61715.png
To apply VFI or super-resolution to any window, first change "Auto Scale" in the Capture section from "Off" to "Windowed Scaling" or "Fullscreen Scaling"
Select the target window, e.g., a browser or media player. Scaling of Browsers currently do not support DRM-protected content. Scaling full-window games (such as Zenless Zone Zero) is not supported either.
Set capture method to default "Graphics Capture". If stuttering occurs, try switching to "GDI".
Set scaling mode to default "Lanczos". For AI super-resolution, select "ONNX" then choose a lightweight 2x super-resolution model from the dropdown (e.g., .onnx models with "SuperUltra" in the name)
For display strategy, if you want the window size to remain unchanged (e.g., frame generation only), select "Don't resize". Otherwise, select "Resize by Ratio" to enlarge the window proportionally.
For scaling strategy, when using methods with significant visual changes (e.g., AI super-resolution), it is recommended to keep the default "Scale and Downscale"
Check "Enable Frame Generation" for smoother window performance
When using frame generation for the first time, be sure to select the ABME algorithm. It is recommended to match the capture framerate with the playback/game framerate — e.g., 24 FPS for anime, minimum 60 FPS for games with locked framerates. By configuring different combinations of frame generation multipliers and capture framerates, you can achieve various target framerates. Ideally, set it to an integer multiple of your monitor's refresh rate for vsync.
For better frame generation quality, consider selecting RIFE. For AMD GPUs or NVIDIA GPUs where you want to avoid compilation/preprocessing time, set the inference backend to DirectML. For NVIDIA GPUs, you can select TensorRT, but the model will undergo pre-compilation first, requiring a wait of approximately 10–30 minutes (the progress bar at the top will show "Preprocessing"). Once completed, you will see the frame generation overlay window. Note that RIFE models will recompile once at resolutions above 2K. Generally, rife 4.6 and rife 4.8 are recommended. Devices below the RTX 40 series should choose rife 4.6, otherwise performance may suffer.
upload log/realtime_pyhook.log if the window still can not be scaled at software's internal forum to help us diagnose the problem.
Interested users, please navigate to SVFI library page → Properties → Betas tab, and select "beta" from the dropdown to join the open beta.
Added new "Tariff" VFI models for NVIDIA GPUs: neu2_nb202 (fast model) and neu2_upg (quality model). The latter offers quality comparable to pg104 with over 2x faster speed! AMD users can select mnn_tariff in the algorithm dropdown (only supports nb202, slower). Recommended for 1080p resolution only; for higher resolutions, set optical flow scale below 0.5
All TensorRT engines require re-precompilation due to model naming convention changes
New Features
Added selected tasks deletion button in main interface
Added GPU selection support for Turbo Mode
Added FFV1 video encoder support
Optimizations and Changes
Removed upscaling ratio restrictions for image super-resolution models
Removed restriction preventing concurrent use of GMFSS and super-resolution
Discontinued INT8 quantization support for super-resolution
Changed deinterlacing filter from yadif to bwdif
Removed EMA VFI model
Removed Mmss VFI model
Removed 3x super-resolution models
Other optimizations
Fixes
Fixed cases where super-resolution output resolution didn't match settings
Fixed playback issues for CPU-encoded HEVC videos on Apple devices
Significantly increases the speed of certain NVIDIA-specific models, located under Advanced Settings
RIFE 4.26 VFI model accelerated by 210%+
GmfSs pg104 accelerated by 15%+
DRBA model for VFI of IRL and conservative Anime accelerated by 300%+
Models under the Super Resolution TensorRT algorithm section (with .onnx extensions) receive a general speed boost
Requires purchasing the Pro version; this option may be incompatible with certain settings
New Features
Supports dragging the error.log file directly into SVFI to load settings
Optimizations and Changes
The surveillance folder can now automatically detect and add newly introduced videos as new tasks, regardless of their creation time
Improved the display of vertical video posters within the software
Silent Mode now only affects the popup that appears after a task is completed. During task startup, a confirmation window will always appear to prevent unreasonable configurations
Allows the computer screen to turn off while the software is running
The official_trt series VFI models have been replaced by Turbo Mode
Other optimizations
Fixes
Fixed the finishing operations after task completion so they now update in real time according to preferences
Fixed an issue where custom output filenames could sometimes cause errors
Added a new feature as Surveillance folders, automatically processing tasks for video files added to target folders based on their creation time. (Hold Shift and drag a folder to quickly set it as a monitored folder)
Added a 150% super-resolution preset, which may lead to changes in some saved preset settings
Optimizations
DRBA now supports arbitrary frame rates
Prevents system sleep during task execution
Improved handling of network fluctuations during task queue execution that could cause queue interruptions
Other optimizations
Fixes
Fixed an issue where using dynamic optical flow could cause errors in some cases
Fixed an issue where using lossless encoding could cause errors in some cases
Fixed an issue where customizing output filenames could cause errors in some cases
The GMFSS frame interpolation model now includes a DRBA frame interpolation mode, which only interpolates the background and not the foreground (characters), catering to users with this specific need
Added the RIFE 4.20+ series to the frame interpolation models
Added a custom output file name style option in Preferences
Added an option in Preferences to whitelist the software in Windows Defender
Added AV1 encoding support for the NVENC encoder in encoding settings
Introduced a new SVFI leaderboard
Added support for transparent channel output in encoding settings, with the community version DLC launching on August 23, 2024
Added super-resolution support for AMD GPUs based on DirectML, allowing AMD GPU users to utilize ONNX models under the TensorRT category
Introduced RTX SR for super-resolution, available only for NVIDIA GPU users, which requires the NVIDIA APP downloaded
Optimizations
Improved the ability to intelligently load super-resolution models supported by architectures from OpenModelDB
Additional optimizations
Fixes
Fixed an issue where high-precision workflows could lose accuracy in certain situations
Resolved an issue with the NVENC encoder where AV1 encoding had an excessively high bitrate
Compared to the official version, the content currently in public beta is as follows:
Image sequence with a transparent channel and video support. Note that the image sequence needs to be saved as png or tiff, and the video will be forced to be exported using Apple ProRes 4444 yuva444p12le to retain the transparent channel; the speed of super-resolution and frame interpolation tasks will be halved because the transparent channel is treated as an independent video stream for processing.
NVIDIA RTX one-click HDR and fast super-resolution model. The NVIDIA APP needs to be installed.
SAIN frame interpolation model
RIFE 4.18 frame interpolation model
Problems that have been attempted to be fixed but are still in public beta for verification:
When the super-resolution ratio (e.g., 2x) does not match the model ratio (e.g., 4x), the output has a mosaic screen.
The output screen of RIFE TensorRT has jagged edges.
Interested friends, please go to the SVFI library page - Properties - Beta section, and pull down "beta" to enter the public beta.
Added marker box for defining the cropping range in the player
Added a search box for the super-resolution model
Added AV1 support for QSVEncC encoder
Added a feature to keep the SVFI window always on top to avoid potential performance degradation due to Windows 11 scheduling, available in preferences
Added an automatic task restart mechanism to continue tasks after errors
Optimizations
Optimized the efficiency of regular super-resolution inference for NVIDIA 40 series graphics cards
A more streamlined Advanced Settings interface. The order of options from left to right remains unchanged. To use the previous advanced settings layout, check Preferences - Classic Advanced Settings.
{STEAM_CLAN_IMAGE}/40952458/117fa697ae6feac26b4c9e9688f43df71733c0c4.png
{STEAM_CLAN_IMAGE}/40952458/f1da94633ea825f482a82ec8403a4b6cdc7e25b7.png
The task input list now includes a Unknown Color Space label. This label indicates that the input video does not have a known color transfer characteristic to assist SVFI in correcting color shift. Manual adjustments are required in the Output Settings under Advanced Settings, specifically in Custom Color Transfer. Generally, bt709 is the default for most 1080p videos.
{STEAM_CLAN_IMAGE}/40952458/ec948f14aac15d92b993a95f453fa823b65a7684.png
{STEAM_CLAN_IMAGE}/40952458/59d746d72a7355f7d182a0b685cd37cadb27f42d.png
Fixes
Fixed an issue where the UI crashes in some cases after terminating a task.
Fixed an issue where video previews would not play correctly in some cases.
Fixed an issue with the TensorRT super-resolution module with tiling reporting errors in some cases.
Fixed an issue where constant frame rate mp4 files were merged into variable frame rate files in some cases.
Fixed an issue in multi-threading mode where it was not possible to specify the super-resolution working graphics card.
Add New Smoothness Optimization method Diff Smooth for incorrectly exported footage (like exporting an original 24fps clip to 60fps with duplicate frames).
Add Scene Extrapolation to reduce duplicate frames in VFI due to scene detection. Powerful VFI models like GMFSS pg 104 are recommended to avoid generating artifacts.
Add new RIFE models with NCNN support.
(Pro) Add alpha(sr intensity) support for SR CUGAN TensorRT models.
New application interface, newly designed preferences page, supporting custom theme colors and custom background images
{STEAM_CLAN_IMAGE}/40952458/b4d7c5855ab2141e45f8bb7b53c6a6ee188657f1.png
{STEAM_CLAN_IMAGE}/40952458/96fe1d3a06021cdbe6f87b54614f77ebcd5e8421.png
Player preview and VFI mask drawing support, area covered by the mask will not be processed in VFI (keep static), suitable for games with HUD (crosshair, control panel, etc.) VFI task.
{STEAM_CLAN_IMAGE}/40952458/a4a46134c97fa67ab0351c59cb981daf4fbe7667.png
{STEAM_CLAN_IMAGE}/40952458/10f0b76113a75608d8cc21869eb7a8223ddbb872.png
{STEAM_CLAN_IMAGE}/40952458/3f3bcac71f9f49ee8d250cbfa1fe5d54d9c289c8.png
PRO: Fast models exclusively for NVIDIA card CUDA core acceleration of TensorRT for CUGAN, RealESRGAN super-resolution models. Choose cugan, realesr under the TensorRT column of super-resolution model selection to use. Changing super-resolution settings requires compilation for these models, please wait patiently for 15 to 20 minutes to achieve more than double acceleration. The new model settings correspond to the old model settings, it's recommended to use the previously well-adjusted super-resolution settings. It's recommended to use graphics cards with 6GB or more video memory. Please make good use of the preset function.
PRO: TensorRT now supports 8-bit quantization of related models, which can further improve speed, recommended for super-resolution models. It's advised to first compile non-int8 trt accelerated models before trying under int8 mode.
{STEAM_CLAN_IMAGE}/40952458/cd46b193f2b416bb5136840a095d86baf4f2c9e0.png
Old version of SVFI 3.23.1 has been uploaded to test branch revision-8. If anyone insists using old version, please go to the SVFI library page - Properties - Test branch dropdown, select "revision-8" to enter the old version of SVFI. Please do remind to add SVFI installation directory to whitelist of your anti-virus software, and remove the files marked as malware from sandbox in anti-virus settings.
New application interface, newly designed preferences page, supporting custom theme colors and custom background images
{STEAM_CLAN_IMAGE}/40952458/b4d7c5855ab2141e45f8bb7b53c6a6ee188657f1.png
{STEAM_CLAN_IMAGE}/40952458/96fe1d3a06021cdbe6f87b54614f77ebcd5e8421.png
Player preview and VFI mask drawing support, area covered by the mask will not be processed in VFI (keep static), suitable for games with HUD (crosshair, control panel, etc.) VFI task.
{STEAM_CLAN_IMAGE}/40952458/a4a46134c97fa67ab0351c59cb981daf4fbe7667.png
{STEAM_CLAN_IMAGE}/40952458/10f0b76113a75608d8cc21869eb7a8223ddbb872.png
{STEAM_CLAN_IMAGE}/40952458/3f3bcac71f9f49ee8d250cbfa1fe5d54d9c289c8.png
Fast models exclusively for NVIDIA card CUDA core acceleration of TensorRT for CUGAN, RealESRGAN, and BasicVSR++ Track3 (T3) super-resolution models. Choose cugan, realesr, or BasicVSRPlusPlusRestore under the TensorRT column of super-resolution model selection, and basicvsrpp_ntire_t3_decompress_max_4x_trt under it to use. Changing super-resolution settings requires compilation for these models, please wait patiently for 15 to 20 minutes to achieve more than double acceleration. The new model settings correspond to the old model settings, it's recommended to use the previously well-adjusted super-resolution settings. It's recommended to use graphics cards with 6GB or more video memory. Please make good use of the preset function.
TensorRT now supports 8-bit quantization of related models, which can further improve speed, recommended for super-resolution models. It's advised to first compile non-int8 trt accelerated models before trying under int8 mode.
{STEAM_CLAN_IMAGE}/40952458/cd46b193f2b416bb5136840a095d86baf4f2c9e0.png
RIFE 4.6, RIFE 4.7 VFI models accelerated by TensorRT
{STEAM_CLAN_IMAGE}/40952458/c97b3acc7147292868d85f67565eb9f0c368f92d.png
Interested buddies, please go to the SVFI library page - Properties - Test branch dropdown, select "beta" to enter the new version of SVFI.
Dear users, due to unforeseen circumstances recently, some users have experienced login issues with our software. We apologize for any inconvenience caused.
To address this, we have added new servers, which have been enabled in the most recent update. Please restart Steam to obtain this update, after which you should be able to log in normally.
Thank you for your understanding and support!
Super-resolution models:BasicVSR++, FTVSR, two super-resolution algorithms with highest academic score, only available with NVIDIA GPUs and the free optimization DLC. These two algorithms support input of multiple images for super-resolution for better detail reconstruction, using the new “input seq-length” option in SR settings. VRAM is heavily consumed. Thus only NVIDIA GPUs with 6G+ VRAM are recommended. Use the tiling function in SR settings appropriately to reduce VRAM usage.
A TensorRT workflow designed for CUGAN , RealESRGAN and BasicVSR++ Track3 that can greatly accelerate (200%+ in speed!) the Super Resolution process, using NVIDIA GPUs with CUDA cores. Available by selecting cugan_trt , RealESR_trt or basicvsrpp_ntire_t3_decompress_max_4x_trt. Ahead compilation is needed, please wait patiently for around 15 to 20 minutes before the actual SR takes off. Previous super resolution settings are recommended to avoid redundant trial and errors. Re-compilation is needed every time the settings are changed. Please make good use of the preset system.
SDR2HDR AI model: FMNet model, generate HDR10 content from SDR in one click.
TensorRT Accelerated RIFE 4.6, GMFSS union_v model
Based on the new decoding process of VSPipe, enable the "Use VSPipe" option in "Output Quality" settings to use QTGMC deinterlacing, fast noise addition and other functions.
Select "beta" branch at Steam-SVFI settings to join Beta now!
About
Steam news for SVFI
This page aggregates Steam news feeds, patch notes, and developer announcements for SVFI, sourced from official Steam community posts.
Major updates, balance changes, and seasonal events often correlate with player-count spikes — cross-reference announcements here with the live charts on the main SVFI statistics page.
Articles link back to Steam for full changelogs. SteamScope refreshes news entries as new posts are published to the game's Steam hub.
FAQ
Frequently asked questions
Where can I read SVFI patch notes?
This page lists official Steam announcements for SVFI, including the latest post "SVFI 8.0". Each entry links to the full changelog on Steam.
How many people are playing SVFI right now?
SVFI currently shows 81 concurrent Steam players on SteamScope. Player counts refresh on a rolling schedule — typically about every 30 minutes for high-priority titles.
How often are SVFI patch notes updated on SteamScope?
SteamScope ingests new Steam announcements as they are published. Check back after major updates, seasons, or balance patches to see fresh entries at the top of this feed.
Is SVFI patch note data official?
SteamScope republishes developer announcements sourced from Steam's public news feeds. SteamScope is not affiliated with Valve Corporation. Always verify critical gameplay changes on the official Steam announcement.
Explore SteamScope
Browse related Steam game intelligence pages and categories.
