Image restoration models

Justin626289/10/26About 4 min

These models appear in the super-resolution list or as restore toggles in Advanced settings. They run at 1× (no upscale) for decompression, watermark removal, debanding, and demosaic. All of them are provided in the Professional version.

Anime IRL Anti-Subtitle Decompress Turbo-Only

Algorithm overview

Algorithm	Edition	Tags	Rec.	AMD
BasicVSR++ T3	Pro	IRL Decompress	★★★★☆	×
Gloom	Pro Beta	Anime IRL	★★★★☆	×
InPaint STTN	Pro Beta	Anti-Subtitle	★★★☆☆	×
LaMa	Pro Beta	Anti-Subtitle	★★★☆☆	×
DeMosaic	Pro Beta	Anime IRL	★★★☆☆	×
DLSS NR	Pro Beta	IRL Turbo-Only	★★★☆☆	×
DeepDeband	Pro Beta	Anime IRL	★★★☆☆	×
FMNet	Pro Beta	Anime IRL	★★★☆☆	×

Warning

Multi-frame restore (T3, Gloom, InPaint) is VRAM-heavy; 6 GB or more is safer.
DeepDeband / FMNet are available only on the public-beta branch.

BasicVSRPlusPlusRestore

IRL-oriented; quality depends on sequence length. T3 is included in the release Professional version; the _trt variant is available only on the public-beta branch.

Model	Edition	Tags	Rec.	Strengths	Weaknesses
basicvsrpp_ntire_t3_decompress_max_4x	Pro	IRL Decompress	★★★★☆	Excellent on heavily compressed footage	Slow and high VRAM; despite the 4x name, SVFI uses it as a 1× restore model
basicvsrpp_ntire_t3_decompress_max_4x_trt	Pro Beta	IRL Decompress	★★☆☆☆	TensorRT attempt	Recommended only for GPUs with at least 16 GB VRAM

Visual comparison

BasicVSR++ Track3 (T3)

basicvsrpp-ntire-t3-decompress-max-4x 1× restore, excellent on heavy compression (slow)

Gloom

Gloom is available only on the public-beta branch. It is a multi-frame enhancement model for low-resolution footage and can restore before upscaling; then specify 2× output resolution to use it as an SR model. Gloom-pro sequence length must be 4n+1 (1 / 5 / 9 / …); try 16 for Gloom and 33 for pro.

Model	Edition	Tags	Rec.	Strengths	Weaknesses
Gloom_x1	Pro Beta	Anime IRL	★★★★☆	Uses multiple frames to enhance the image	High VRAM
Gloom-pro_x1	Pro Beta	Anime IRL	★★★★☆	Stronger pro path	At least 8 GB VRAM recommended

InPaint Watermark Removal Model

InPaint_STTN is available only in the public-beta Professional version. It performs 1× restoration and needs a static mask from the internal player.

Model	Edition	Tags	Rec.	Strengths	Weaknesses
InPaint_STTN	Pro Beta	Anti-Subtitle	★★★☆☆	Static watermarks inside the mask can be removed automatically	Fails on solid/static backgrounds; the mask must contain motion

Steps

Enable super-resolution and select this model

Open the internal player

Enable the mask

Draw and save

Leave enough moving content inside the mask, or detection fails.

Warning

Watermarks on solid / fully static backgrounds inpaint poorly.

Click Encode. Sequence length greater than 30 is recommended.

InPaint watermark removal demo

Sequence length greater than 30 is recommended

LaMa

LaMa is a classic image erasure model. SVFI automatically detects subtitles in video and removes them. It is not suitable for solid-color still backgrounds.

Model	Edition	Tags	Rec.	Strengths	Weaknesses
big-lama	Pro Beta	Anti-Subtitle	★★★☆☆	Classic erase network; works with a mask	Also weak on solid stills

DeMosaic

DeMosaic removes mosaic and pixelation from special-source footage. SVFI automatically detects and removes mosaics. In debug mode, it does not remove them, but shows the detected regions so the mosaic-detection threshold can be adjusted.

Model	Edition	Tags	Rec.	Strengths	Weaknesses
bvpp_x1 (DeMosaic, LADA)	Pro Beta	Anime IRL	★★★☆☆	Removes mosaics	Incompatible with the 16-bit workflow

DLSS NR

DLSS NR is a 1× denoise model that must use Turbo and currently supports only the 8-bit workflow. It is designed to add realistic lighting detail to game recordings.

Model	Edition	Tags	Rec.	Strengths	Weaknesses
DLSSNR_x1	Pro Beta	IRL Turbo-Only	★★★☆☆	Adds realistic lighting detail	Turbo required; currently 8-bit only

DLSS NR demo

Enable reconstruction guidance (reconstructed input motion vectors and depth information)

DeepDeband

DeepDeband uses an AI algorithm to remove banding from video. Available only on the public-beta branch.

Model	Edition	Tags	Rec.	Strengths	Weaknesses
deepdeband-f / deepdeband-w	Pro Beta	Anime IRL	★★★☆☆	Reduces color banding	Unavailable on Turbo

FMNet

FMNet uses an AI algorithm to convert SDR video to HDR10 video. Available only on the public-beta branch. When enabled, the video becomes HDR10 with static metadata.

Model	Edition	Tags	Rec.	Strengths	Weaknesses
FMNet anime_v1 / anime_v2 / final	Pro Beta	Anime IRL	★★★☆☆	Restore for HDR and color-grading scenarios	Unavailable on Turbo