Skip to content

Commit df0add1

Browse files
Kim2091Kim2091
authored
Add new models (#475)
* Add new models * Update architectures.json * fix db * fix? * fix again
1 parent f3ff47c commit df0add1

17 files changed

+709
-2
lines changed

‎data/architectures.json‎

Lines changed: 16 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -145,6 +145,14 @@
145145
"pytorch"
146146
]
147147
},
148+
"rcan": {
149+
"name": "RCAN",
150+
"input": "image",
151+
"compatiblePlatforms": [
152+
"pytorch",
153+
"onnx"
154+
]
155+
},
148156
"real-cugan": {
149157
"name": "Real-CUGAN",
150158
"input": "image",
@@ -236,5 +244,13 @@
236244
"pytorch",
237245
"onnx"
238246
]
247+
},
248+
"tscunet": {
249+
"name": "TSCUNet",
250+
"input": "video",
251+
"compatiblePlatforms": [
252+
"pytorch",
253+
"onnx"
254+
]
239255
}
240256
}

‎data/collections.json‎

Lines changed: 30 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -10,6 +10,24 @@
1010
"2x-AnimeSharpV2-MoSR-Soft"
1111
]
1212
},
13+
"c-animesharpv3": {
14+
"name": "AnimeSharpV3",
15+
"author": "kim2091",
16+
"description": "",
17+
"models": [
18+
"2x-AnimeSharpV3",
19+
"2x-AnimeSharpV3-RCAN"
20+
]
21+
},
22+
"c-animesharpv4": {
23+
"name": "AnimeSharpV4",
24+
"author": "kim2091",
25+
"description": "",
26+
"models": [
27+
"2x-AnimeSharpV4",
28+
"2x-AnimeSharpV4-Fast-RCAN-PU"
29+
]
30+
},
1331
"c-normal-map-upscaling": {
1432
"name": "Normal Map Upscaling",
1533
"description": "This collection contain my RG0 normal map upscaling models.\n\nAll models here are for upscaling *tangent-space* normal maps in RG0 format. RG0 means that the B channel is set to 0. These models will work not correctly if you give them images with non-zero B channel, so you either have to zero the B channel manually or use tool like chaiNNer to do it.\n\n## DDS Compression\n\nI made 3 versions: \n- Normal RG0 is for uncompressed normal map textures. Since it hasn't been trained on compression artifacts, it's highly sensitive to quantization artifacts and noise.\n- Normal RG0 BC1 is for BC1-compressed DDS normal map textures.\n- Normal RG0 BC7 is for BC7-compressed DDS normal map textures. This model sometimes produces images that aren't as sharp. In those cases, you can try the BC1 version to see whether it gives better results.",
@@ -30,7 +48,18 @@
3048
"4x-PBRify-UpscalerSPANV4",
3149
"4x-PBRify-UpscalerSIR-M-V2",
3250
"4x-PBRify-UpscalerDAT2-V1",
33-
"4x-PBRify-RPLKSRd-V3"
51+
"4x-PBRify-RPLKSRd-V3",
52+
"4x-PBRify-UpscalerV4"
53+
],
54+
"author": "kim2091"
55+
},
56+
"c-gameup": {
57+
"name": "GameUp",
58+
"description": "GameUp is a set of video upscaling models intended for upscaling and restoring video game footage",
59+
"models": [
60+
"2x-GameUpV2-TSCUNet",
61+
"2x-GameUpV2-TSCUNet-Small",
62+
"2x-GameUp-TSCUNet"
3463
],
3564
"author": "kim2091"
3665
}

‎data/models/1x-BroadcastToStudio-Compact.json‎

Lines changed: 53 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,53 @@
1+
{
2+
"name": "BroadcastToStudio Compact",
3+
"author": "kim2091",
4+
"license": "CC-BY-NC-SA-4.0",
5+
"tags": [
6+
"cartoon",
7+
"compression-removal",
8+
"deblur",
9+
"restoration"
10+
],
11+
"description": "Purpose: Cartoons\n\nThis is a simple retrain of SaurusX's 1x_BroadcastToStudioLite_485k model from a couple years ago. This one is trained on compact, actually has less artifacts, and is significantly faster.\n\n__Comparisons:__ <https://slow.pics/c/oGwHyYym>\n![1733071729 5752351](https://github.com/user-attachments/assets/f247b870-d49e-4c37-ad39-c4565efe164d)",
12+
"date": "2024-12-01",
13+
"architecture": "compact",
14+
"size": [
15+
"64nf",
16+
"16nc"
17+
],
18+
"scale": 1,
19+
"inputChannels": 3,
20+
"outputChannels": 3,
21+
"resources": [
22+
{
23+
"platform": "pytorch",
24+
"type": "pth",
25+
"size": 2400484,
26+
"sha256": "f4876edc5f12783395c444bbb1d2f1bc304ed69b2cc82409dff9d6dad6fbd596",
27+
"urls": [
28+
"https://github.com/Kim2091/Kim2091-Models/releases/download/1x-BroadcastToStudio_Compact/1x-BroadcastToStudio_Compact.pth"
29+
]
30+
},
31+
{
32+
"platform": "onnx",
33+
"type": "onnx",
34+
"size": 1200682,
35+
"sha256": "52836c782140058bcc695e90102c3ef54961ebab2c12e66298eaba25d42570bc",
36+
"urls": [
37+
"https://github.com/Kim2091/Kim2091-Models/releases/download/1x-BroadcastToStudio_Compact/1x-BroadcastToStudio_Compact-fp16.onnx"
38+
]
39+
}
40+
],
41+
"trainingIterations": 25000,
42+
"trainingBatchSize": 8,
43+
"dataset": "BroadcastToStudio",
44+
"datasetSize": 6000,
45+
"pretrainedModelG": "1x-SwatKats-Compact",
46+
"images": [
47+
{
48+
"type": "paired",
49+
"LR": "https://i.slow.pics/DBrq6k0g.webp",
50+
"SR": "https://i.slow.pics/XJmeT86k.webp"
51+
}
52+
]
53+
}

‎data/models/1x-SuperScale-Alt-RPLKSR-S.json‎

Lines changed: 41 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,41 @@
1+
{
2+
"name": "SuperScale_Alt_RPLKSR_S",
3+
"author": "kim2091",
4+
"license": "CC-BY-NC-SA-4.0",
5+
"tags": [
6+
"anti-aliasing",
7+
"game-screenshots",
8+
"photo",
9+
"restoration"
10+
],
11+
"description": "Purpose: Anti-aliasing, Restoration\n\nI was bored, so I did this. This model uses DPID as the scaling algorithm for the HRs. The original images were 8k or 12k. It's significantly sharper than Box/Area scaling, yet does a great job with aliasing. This allows for a very sharp model with minimal artifacts, even on the SPAN version.\n\nThe main model is trained on 12k images captured with Nvidia Ansel. It took about 2 days capturing manual 4k and 12k pairs for this model. The 4k captures were used as the LR, the 12k captures were resized to 4k with DPID with randomized lambda values, then trained on as HRs. \n\nThe Alt model is trained exclusively on 8k images from my 8k dataset, resized to 4k with dpid. This provides a clearer result with less noise, but it doesn't handle long edges well at all.\n\nThanks to CF2lter for advice on preparing the dataset, and umzi2 for creating the [rust version of DPID](<https://github.com/umzi2/pepedpid>). \n\n**Showcase:** https://slow.pics/c/TCyqje9K\n![Animation (2)](https://github.com/user-attachments/assets/fb9f010a-a4e1-4537-8b23-9a69974011c6)",
12+
"date": "2025-05-05",
13+
"architecture": "realplksr",
14+
"size": [
15+
"Small",
16+
"Tiny"
17+
],
18+
"scale": 1,
19+
"inputChannels": 3,
20+
"outputChannels": 3,
21+
"resources": [
22+
{
23+
"platform": "pytorch",
24+
"type": "safetensors",
25+
"size": 1195070,
26+
"sha256": "f5844dd72922a6579cf73c44de5cf35f0d700bd407982ae8d03bcdf720924425",
27+
"urls": [
28+
"https://github.com/Kim2091/Kim2091-Models/releases/download/1x-SuperScale/1x-SuperScale_Alt_RPLKSR_S.safetensors"
29+
]
30+
}
31+
],
32+
"dataset": "8k Dataset V3, Custom Ansel dataset",
33+
"images": [
34+
{
35+
"type": "paired",
36+
"caption": "1",
37+
"LR": "https://i.slow.pics/543aXqwG.webp",
38+
"SR": "https://i.slow.pics/o3fwZmcB.webp"
39+
}
40+
]
41+
}

‎data/models/1x-SuperScale-RPLKSR-S.json‎

Lines changed: 38 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,38 @@
1+
{
2+
"name": "SuperScale_RPLKSR_S",
3+
"author": "kim2091",
4+
"license": "CC-BY-NC-SA-4.0",
5+
"tags": [
6+
"anti-aliasing",
7+
"game-screenshots",
8+
"photo",
9+
"restoration"
10+
],
11+
"description": "Purpose: Anti-aliasing, Restoration\n\nI was bored, so I did this. This model uses DPID as the scaling algorithm for the HRs. The original images were 8k or 12k. It's significantly sharper than Box/Area scaling, yet does a great job with aliasing. This allows for a very sharp model with minimal artifacts, even on the SPAN version.\n\nThe main model is trained on 12k images captured with Nvidia Ansel. It took about 2 days capturing manual 4k and 12k pairs for this model. The 4k captures were used as the LR, the 12k captures were resized to 4k with DPID with randomized lambda values, then trained on as HRs. \n\nThe Alt model is trained exclusively on 8k images from my 8k dataset, resized to 4k with dpid. This provides a clearer result with less noise, but it doesn't handle long edges well at all.\n\nThanks to CF2lter for advice on preparing the dataset, and umzi2 for creating the [rust version of DPID](<https://github.com/umzi2/pepedpid>). \n\n**Showcase:** https://slow.pics/c/TCyqje9K\n![Animation (2)](https://github.com/user-attachments/assets/fb9f010a-a4e1-4537-8b23-9a69974011c6)",
12+
"date": "2025-05-05",
13+
"architecture": "realplksr",
14+
"size": null,
15+
"scale": 1,
16+
"inputChannels": 3,
17+
"outputChannels": 3,
18+
"resources": [
19+
{
20+
"platform": "pytorch",
21+
"type": "safetensors",
22+
"size": 1195070,
23+
"sha256": "f5844dd72922a6579cf73c44de5cf35f0d700bd407982ae8d03bcdf720924425",
24+
"urls": [
25+
"https://github.com/Kim2091/Kim2091-Models/releases/download/1x-SuperScale/1x-SuperScale_RPLKSR_S.safetensors"
26+
]
27+
}
28+
],
29+
"dataset": "8k Dataset V3, Custom Ansel dataset",
30+
"images": [
31+
{
32+
"type": "paired",
33+
"caption": "1",
34+
"LR": "https://i.slow.pics/543aXqwG.webp",
35+
"SR": "https://i.slow.pics/V1xP5Zh5.webp"
36+
}
37+
]
38+
}

‎data/models/1x-SuperScale.json‎

Lines changed: 38 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,38 @@
1+
{
2+
"name": "SuperScale",
3+
"author": "kim2091",
4+
"license": "CC-BY-NC-SA-4.0",
5+
"tags": [
6+
"anti-aliasing",
7+
"game-screenshots",
8+
"photo",
9+
"restoration"
10+
],
11+
"description": "Purpose: Anti-aliasing, Restoration\n\nI was bored, so I did this. This model uses DPID as the scaling algorithm for the HRs. The original images were 8k or 12k. It's significantly sharper than Box/Area scaling, yet does a great job with aliasing. This allows for a very sharp model with minimal artifacts, even on the SPAN version.\n\nThe main model is trained on 12k images captured with Nvidia Ansel. It took about 2 days capturing manual 4k and 12k pairs for this model. The 4k captures were used as the LR, the 12k captures were resized to 4k with DPID with randomized lambda values, then trained on as HRs. \n\nThe Alt model is trained exclusively on 8k images from my 8k dataset, resized to 4k with dpid. This provides a clearer result with less noise, but it doesn't handle long edges well at all.\n\nThanks to CF2lter for advice on preparing the dataset, and umzi2 for creating the [rust version of DPID](<https://github.com/umzi2/pepedpid>). \n\n**Showcase:** https://slow.pics/c/TCyqje9K\n![Animation (2)](https://github.com/user-attachments/assets/fb9f010a-a4e1-4537-8b23-9a69974011c6)",
12+
"date": "2025-05-05",
13+
"architecture": "span",
14+
"size": null,
15+
"scale": 1,
16+
"inputChannels": 3,
17+
"outputChannels": 3,
18+
"resources": [
19+
{
20+
"platform": "pytorch",
21+
"type": "safetensors",
22+
"size": 1195070,
23+
"sha256": "f5844dd72922a6579cf73c44de5cf35f0d700bd407982ae8d03bcdf720924425",
24+
"urls": [
25+
"https://github.com/Kim2091/Kim2091-Models/releases/download/1x-SuperScale/1x-SuperScale_SPAN.safetensors"
26+
]
27+
}
28+
],
29+
"dataset": "8k Dataset V3, Custom Ansel dataset",
30+
"images": [
31+
{
32+
"type": "paired",
33+
"caption": "1",
34+
"LR": "https://i.slow.pics/543aXqwG.webp",
35+
"SR": "https://i.slow.pics/pKA297Fp.webp"
36+
}
37+
]
38+
}

‎data/models/1x-UnResizeOnly-RCAN.json‎

Lines changed: 38 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,38 @@
1+
{
2+
"name": "UnResizeOnly_RCAN",
3+
"author": "kim2091",
4+
"license": "CC-BY-NC-SA-4.0",
5+
"tags": [
6+
"anti-aliasing",
7+
"restoration"
8+
],
9+
"description": "Purpose: Artifact Removal\n\nA version of UnResize trained on RCAN, which is faster and provides better quality than ESRGAN\n\nThis model does **not remove compression or perform deblurring**, unlike the original UnResize models. __It **only** removes scaling artifacts.__\n\nI've attached the script I used to create the dataset (it utilizes imagemagick) and the config for [traiNNer-redux](https://github.com/the-database/traiNNer-redux)",
10+
"date": "2025-01-06",
11+
"architecture": "rcan",
12+
"size": null,
13+
"scale": 1,
14+
"inputChannels": 3,
15+
"outputChannels": 3,
16+
"resources": [
17+
{
18+
"platform": "pytorch",
19+
"type": "safetensors",
20+
"size": 30757598,
21+
"sha256": "b88289e770207e634181b595845e0d240cc397714aefe659bcf9b70478b64373",
22+
"urls": [
23+
"https://github.com/Kim2091/Kim2091-Models/releases/download/1x-UnResizeOnly_RCAN/1x-UnResizeOnly_RCAN.pth"
24+
]
25+
}
26+
],
27+
"trainingIterations": 95000,
28+
"trainingBatchSize": 8,
29+
"dataset": "UltraSharpV2_Ethical, DigitalArtV3, ModernAnimation1080_v3, Kim2091's 8k Dataset V2",
30+
"datasetSize": 13000,
31+
"images": [
32+
{
33+
"type": "paired",
34+
"LR": "https://imgsli.com/i/1cdd6669-3798-4359-b62d-0a18a38a810b.jpg",
35+
"SR": "https://imgsli.com/i/aaa71ef9-e7c7-4af1-8aee-d653b903bd6f.jpg"
36+
}
37+
]
38+
}

‎data/models/2x-AnimeSharpV3-RCAN.json‎

Lines changed: 45 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,45 @@
1+
{
2+
"name": "AnimeSharpV3_RCAN",
3+
"author": "kim2091",
4+
"license": "CC-BY-NC-SA-4.0",
5+
"tags": [
6+
"anime",
7+
"cartoon"
8+
],
9+
"description": "Purpose: Anime\n\nThis release contains an ESRGAN and an RCAN version. Both provide superior quality compared to AnimeSharpV2 in nearly every scenario. It has most of the advantages of the old V2 Sharp models, while not having issues with depth of field. \n\nThe RCAN model outperforms the ESRGAN model by a significant margin, with much more consistent generation and overall better detail retention. Currently it is NOT compatible with chaiNNer, but will be available on the nightly build soon (hopefully).\n\nRCAN vs ESRGAN: https://slow.pics/c/Zqgl62Ni\n\n__Comparisons:__ <https://slow.pics/c/A2BRSa0U>\n\n\n![1729798851 305732](https://github.com/user-attachments/assets/4d7aaf33-5a39-4f75-8a05-90b90a693e49)",
10+
"date": "2024-10-03",
11+
"architecture": "rcan",
12+
"size": null,
13+
"scale": 2,
14+
"inputChannels": 3,
15+
"outputChannels": 3,
16+
"resources": [
17+
{
18+
"platform": "pytorch",
19+
"type": "safetensors",
20+
"size": 31053198,
21+
"sha256": "9c802d4d40238605d4ae8902f1f170c729ecdde142078838329b490d796292ee",
22+
"urls": [
23+
"https://github.com/Kim2091/Kim2091-Models/releases/download/2x-AnimeSharpV3/2x-AnimeSharpV3_RCAN.safetensors"
24+
]
25+
}
26+
],
27+
"trainingIterations": 140000,
28+
"trainingBatchSize": 8,
29+
"trainingOTF": false,
30+
"dataset": "ModernAnimation1080_v3",
31+
"datasetSize": 3000,
32+
"pretrainedModelG": "4x-ESRGAN",
33+
"images": [
34+
{
35+
"type": "paired",
36+
"LR": "https://i.slow.pics/YAyhw81T.webp",
37+
"SR": "https://i.slow.pics/j8BtouUm.webp"
38+
},
39+
{
40+
"type": "paired",
41+
"LR": "https://i.slow.pics/QFNEusLp.webp",
42+
"SR": "https://i.slow.pics/tLL6gO5w.webp"
43+
}
44+
]
45+
}

‎data/models/2x-AnimeSharpV3.json‎

Lines changed: 56 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,56 @@
1+
{
2+
"name": "AnimeSharpV3",
3+
"author": "kim2091",
4+
"license": "CC-BY-NC-SA-4.0",
5+
"tags": [
6+
"anime",
7+
"cartoon"
8+
],
9+
"description": "Purpose: Anime\n\nThis release contains an ESRGAN and an RCAN version. Both provide superior quality compared to AnimeSharpV2 in nearly every scenario. It has most of the advantages of the old V2 Sharp models, while not having issues with depth of field. \n\nThe RCAN model outperforms the ESRGAN model by a significant margin, with much more consistent generation and overall better detail retention. Currently it is NOT compatible with chaiNNer, but will be available on the nightly build soon (hopefully).\n\nRCAN vs ESRGAN: https://slow.pics/c/Zqgl62Ni\n\n__Comparisons:__ <https://slow.pics/c/A2BRSa0U>\n\n\n![1729798851 305732](https://github.com/user-attachments/assets/4d7aaf33-5a39-4f75-8a05-90b90a693e49)",
10+
"date": "2024-10-03",
11+
"architecture": "esrgan",
12+
"size": [
13+
"64nf",
14+
"23nb"
15+
],
16+
"scale": 2,
17+
"inputChannels": 3,
18+
"outputChannels": 3,
19+
"resources": [
20+
{
21+
"platform": "pytorch",
22+
"type": "pth",
23+
"size": 67104146,
24+
"sha256": "d5722d738002c1353ce4bbc3fb44fe2ecf01606c713cdb5853f772b08af84f53",
25+
"urls": [
26+
"https://github.com/Kim2091/Kim2091-Models/releases/download/2x-AnimeSharpV3/2x-AnimeSharpV3.pth"
27+
]
28+
},
29+
{
30+
"platform": "onnx",
31+
"type": "onnx",
32+
"size": 33619368,
33+
"sha256": "fe4cbe50bfc8b20dfcb16b0935ef4dbdb64547224bee17ec2f496385bc37a71e",
34+
"urls": [
35+
"https://github.com/Kim2091/Kim2091-Models/releases/download/2x-AnimeSharpV3/2x-AnimeSharpV3-fp16.onnx"
36+
]
37+
}
38+
],
39+
"trainingIterations": 140000,
40+
"trainingBatchSize": 8,
41+
"dataset": "ModernAnimation1080_v3",
42+
"datasetSize": 3000,
43+
"pretrainedModelG": "4x-ESRGAN",
44+
"images": [
45+
{
46+
"type": "paired",
47+
"LR": "https://i.slow.pics/TwfO01xU.webp",
48+
"SR": "https://i.slow.pics/4LuILsUF.webp"
49+
},
50+
{
51+
"type": "paired",
52+
"LR": "https://i.slow.pics/PsraU7Kd.webp",
53+
"SR": "https://i.slow.pics/qdXwDmJ6.webp"
54+
}
55+
]
56+
}

0 commit comments

Comments
 (0)