Models

Restoration models

32 restoration models. Each section says what the model is, whether it may be used commercially and how fast it is; licences and checks are under Details.

Input: one RGB image. Output: the restored (upscaled) image, model.scale x the input size, RGB in [0, 1]. tile= runs large images in overlapping tiles (as ComfyUI).

Measured on RTX 5060 Ti, torch 2.14.1+cu130; fastforks bench MODEL --size HxW measures yours. The summary uses one input, 512x512: kernels time with its speed-up over the faster original run, and kernels peak reserved memory with its ratio to the smaller original one. Where measured (one variant per family), a variant's table also lists portable under torch.compile and the fp16 TensorRT engine of portable (export; memory: the engine's device memory + torch's peak). – = not measured; no kernels = the family has no kernels backend (reason under its table). Reserved memory is the allocator's cache: kernels (CUDA graphs, cuDNN workspaces) can reserve a few percent to some tens of percent more than portable at the same allocated peak.

Other devices: Apple M1 Pro 16 GB, NVIDIA L4, NVIDIA L40S, NVIDIA A10, NVIDIA A100-SXM4-40GB, NVIDIA RTX PRO 6000; across devices.

modelcommercialkernels time (ms),
speed-up vs original
kernels peak memory (MiB, reserved),
ratio to original
big-lamayes21.2, 1.8x284, 0.99x
codeformerno40.4, 2.1x496, 0.46x
dat-2-x4yes1221, 4.0x2378, 0.94x
dat-light-x4yes264, 8.3x1472, 0.89x
dat-x4yes1391, 6.9x2384, 0.94x
drct-l-x4yes1399, 10.5x2850, 0.41x
drct-x4yes718, 10.4x2766, 0.7x
gfpgan-v1.4yes18.0, 1.2x530, 0.79x
hat-s-x4yes956, 2.5x1888, 0.33x
hat-x4yes1427, 2.1x2470, 0.4x
nafnet-goproyes55.3, 1.2x550, 0.9x
nafnet-siddyes56.9, 1.3x650, 0.89x
real-hat-gan-x4yes1427, 2.1x2470, 0.4x
realesr-general-x4v3yes24.0, 1.8x164, 0.8x
realesrgan-x2plusyes80.2, 1.6x394, 0.6x
realesrgan-x4plusyes403, 1.6x1434, 0.57x
realesrgan-x4plus-anime-6byes132, 1.6x1410, 0.57x
realplksr-nomoswebphoto-x4yes130, 2.1x216, 0.68x
restormer-debluryes180, 1.3x918, 0.65x
restormer-derainyes180, 1.3x920, 0.65x
restormer-siddyes179, 1.2x940, 0.66x
seedvr2-3byes2296, 5.7x6772, 0.44x
seedvr2-7byes27137034
span-x2yes14.9, 1.6x522, 0.81x
span-x4yes14.6, 1.8x522, 0.81x
swin2sr-x2yes721, 3.6x1636, 1.0x
swin2sr-x4yes745, 3.9x1736, 0.8x
swin2sr-x4-realyes751, 3.4x1734, 0.65x
swinir-l-x4-realyes1036, 3.6x1934, 0.7x
swinir-m-x2yes553, 5.9x1454, 0.8x
swinir-m-x4yes576, 6.1x1456, 0.71x
swinir-m-x4-realyes584, 4.2x1814, 0.61x

All models

Modes, licences, training data and outputs per model

Commercial status per variant: yes if the code and the weights licence both allow commercial use (the training data is listed but does not change it; weights without a licence of their own are assumed to carry the code licence); no if either forbids it; unclear for custom licence terms (treat as no).

modelmodescode licenceweights licencetraining datacommercialoutputs
big-lamaportable, kernelsApache-2.0Apache-2.0places365yesimage
codeformerportable, kernelsLicenseRef-noncommercialLicenseRef-noncommercialffhqno (code: LicenseRef-noncommercial (noncommercial); weights: LicenseRef-noncommercial)image
dat-2-x4portable, kernelsApache-2.0Apache-2.0div2k, flickr2kyesimage
dat-light-x4portable, kernelsApache-2.0Apache-2.0div2k, flickr2kyesimage
dat-x4portable, kernelsApache-2.0Apache-2.0div2k, flickr2kyesimage
drct-l-x4portable, kernelsMITMITimagenet-1k, div2k, flickr2kyesimage
drct-x4portable, kernelsMITMITimagenet-1k, div2k, flickr2kyesimage
gfpgan-v1.4portable, kernelsApache-2.0Apache-2.0ffhqyesimage
hat-s-x4portable, kernelsApache-2.0Apache-2.0div2k, flickr2kyesimage
hat-x4portable, kernelsApache-2.0Apache-2.0imagenet-1k, div2k, flickr2kyesimage
nafnet-goproportable, kernelsMITMITgoproyesimage
nafnet-siddportable, kernelsMITMITsiddyesimage
real-hat-gan-x4portable, kernelsApache-2.0Apache-2.0div2k, flickr2k, ostyesimage
realesr-general-x4v3portable, kernelsBSD-3-ClauseBSD-3-Clausediv2k, flickr2k, ostyesimage
realesrgan-x2plusportable, kernelsBSD-3-ClauseBSD-3-Clausediv2k, flickr2k, ostyesimage
realesrgan-x4plusportable, kernelsBSD-3-ClauseBSD-3-Clausediv2k, flickr2k, ostyesimage
realesrgan-x4plus-anime-6bportable, kernelsBSD-3-ClauseBSD-3-Clausediv2k, flickr2k, ostyesimage
realplksr-nomoswebphoto-x4portable, kernelsApache-2.0CC-BY-4.0nomos-v2yesimage
restormer-deblurportable, kernelsMITMITgoproyesimage
restormer-derainportable, kernelsMITMITrain13kyesimage
restormer-siddportable, kernelsMITMITsiddyesimage
seedvr2-3bportable, kernelsApache-2.0Apache-2.0unstatedyesimage
seedvr2-7bportable, kernelsApache-2.0Apache-2.0unstatedyesimage
span-x2portable, kernelsApache-2.0Apache-2.0div2k, flickr2kyesimage
span-x4portable, kernelsApache-2.0Apache-2.0div2k, flickr2kyesimage
swin2sr-x2portable, kernelsApache-2.0Apache-2.0div2k, flickr2kyesimage
swin2sr-x4portable, kernelsApache-2.0Apache-2.0div2k, flickr2kyesimage
swin2sr-x4-realportable, kernelsApache-2.0Apache-2.0div2k, flickr2kyesimage
swinir-l-x4-realportable, kernelsApache-2.0Apache-2.0div2k, flickr2k, ost, wed, ffhq, manga109, scut-ctw1500yesimage
swinir-m-x2portable, kernelsApache-2.0Apache-2.0div2k, flickr2kyesimage
swinir-m-x4portable, kernelsApache-2.0Apache-2.0div2k, flickr2kyesimage
swinir-m-x4-realportable, kernelsApache-2.0Apache-2.0div2k, flickr2k, ostyesimage

Training datasets