【搬运教程】各显卡对应的模型参数表

Lau9

星级打分

• 1
• 2
• 3
• 4
• 5

平均分:NAN  参与人数:0  我的评分:未评

Adabelief Enabled

GPU​

VRAM​

CPU​

RAM​

Architecture​

Resolution​

AE Dims​

E Dims​

D Dims​

D Mask Dims​

Batch Size​

Iteration Times​

GPU Optimizer​

GTX 1060​

6​

i5-4690K​

16​

LIAE-UD​

320​

264​

72​

72​

24​

5​

6500​

FALSE​

GTX 1080 Ti​

11​

i7-4770K​

16​

DF-UD​

320​

320​

72​

72​

16​

8​

700​

TRUE​

GTX Titan X​

12​

i7-2600K​

24​

LIAE-UDT​

224​

256​

64​

64​

16​

6​

1035​

TRUE​

RTX 2080​

8​

i7-8700K​

32​

LIAE-UD​

256​

256​

64​

64​

22​

8​

500​

FALSE​

RTX 3050​

8​

i5-4670​

8​

DF-UD​

320​

256​

64​

64​

22​

4​

680​

TRUE​

RTX 3060​

12​

i5-8400​

32​

DF-UD​

320​

320​

96​

96​

32​

7​

1350​

TRUE​

RTX 3060​

12​

R5-5600X​

32​

DF-UD​

384​

256​

64​

64​

22​

8​

1140​

TRUE​

RTX 3060​

12​

R5-3600​

32​

DF-UD​

384​

256​

64​

64​

22​

8​

1130​

TRUE​

RTX 3060​

12​

i5-8400​

32​

DF-UD​

320​

360​

90​

90​

22​

7​

1350​

TRUE​

RTX 3060​

12​

i5-8400​

32​

DF-UD​

320​

288​

80​

80​

22​

9​

1350​

TRUE​

RTX 3060​

12​

i5-8400​

32​

DF-UDT​

320​

320​

88​

88​

22​

7​

1350​

TRUE​

RTX 3060​

12​

i5-8400​

32​

DF-UDT​

256​

300​

80​

64​

22​

17​

1350​

TRUE​

RTX 3060​

12​

i5-8400​

32​

DF-UDT​

288​

300​

80​

80​

22​

10​

1350​

TRUE​

RTX 3060​

12​

i5-8400​

32​

DF-UDT​

320​

360​

90​

90​

22​

5​

1350​

TRUE​

RTX 3060​

12​

i5-8400​

32​

LIAE-UDT​

320​

256​

72​

72​

32​

9​

1350​

TRUE​

RTX 3060​

12​

i5-8400​

32​

LIAE-UD​

320​

288​

72​

72​

22​

7​

1350​

TRUE​

RTX 3060​

12​

i5-8400​

32​

LIAE-UD​

256​

256​

64​

64​

22​

18​

1350​

TRUE​

RTX 3060​

12​

i5-8400​

32​

LIAE-UD​

256​

256​

80​

80​

22​

13​

1350​

TRUE​

RTX 3060​

12​

i5-8400​

32​

LIAE-UD​

320​

256​

64​

64​

22​

10​

1350​

TRUE​

RTX 3060​

12​

i5-8400​

32​

LIAE-UDT​

320​

320​

88​

88​

22​

7​

1350​

TRUE​

RTX 3090​

24​

R9-3900X​

32​

DF-UD​

384​

512​

112​

112​

16​

8​

1013​

TRUE​

RTX 3090​

24​

R3-3900X​

32​

DF-UD​

320​

512​

112​

112​

16​

16​

1074​

TRUE​

RTX 3090​

24​

i7-5820K​

48​

DF-UD​

416​

416​

104​

104​

26​

8​

1170​

TRUE​

RTX 3090​

24​

R9-3900X​

32​

LIAE-UDT​

224​

512​

64​

64​

16​

48​

998​

TRUE​

RTX 3090​

24​

R9-3900X​

32​

LIAE-UDT​

288​

352​

128​

128​

16​

16​

1320​

TRUE​

Tesla V100​

16​

Colab​

25​

DF-UD​

384​

352​

88​

88​

16​

8​

1000​

TRUE​

Tesla V100​

16​

Colab​

25​

DF-UD​

320​

416​

104​

104​

16​

8​

850​

TRUE​

Tesla V100​

16​

Colab​

25​

DF-UD​

384​

320​

80​

80​

22​

8​

900​

TRUE​

Tesla V100​

16​

Colab​

25​

DF-UD​

448​

256​

64​

64​

22​

8​

950​

TRUE​

Below are settings with no adabelief, mostly older results with base archis, consider these "legacy" and mostly irrelevant, do not train with adabelief disabled and keep in mind that without -D flag models are much heavier to train (achieving higher resolution is higher, but you could get better quality with something like DF-UT or LIAE-UT over -UD/UDT variants, but it will use a lot of VRAM, most people use -UD/UDT variants these days.

Adabelief Disabled

GPU​

VRAM​

CPU​

RAM​

Architecture​

Resolution​

AE Dims​

E Dims​

D Dims​

D Mask Dims​

Batch Size​

Iteration Times​

GPU Optimizer​

GTX 750 Ti​

2​

i5-4690K​

32​

LIAE​

112​

256​

64​

64​

22​

4​

1450​

FALSE​

GTX 970​

4​

i7-2600K​

12​

DF​

96​

256​

64​

64​

22​

4​

700​

TRUE​

GTX 970​

4​

i7-2600K​

12​

DF​

96​

256​

64​

64​

22​

4​

700​

TRUE​

GTX 1050 Ti​

4​

i5-3570K​

12​

DF​

128​

192​

48​

48​

36​

2​

645​

TRUE​

GTX 1050 Ti​

4​

i5-3570K​

12​

DF​

128​

192​

48​

48​

36​

2​

645​

FALSE​

GTX 1050 Ti​

4​

E5-1620​

16​

LIAE​

128​

128​

80​

48​

16​

4​

520​

TRUE​

GTX 1060​

6​

i5-4670K​

16​

DF​

192​

256​

64​

64​

22​

7​

1200​

FALSE​

GTX 1060​

6​

i5-4590​

16​

DF​

192​

256​

64​

64​

22​

6​

1400​

FALSE​

GTX 1060​

6​

i5-4670K​

16​

DF​

128​

450​

64​

64​

22​

10​

750​

TRUE​

GTX 1060​

6​

i7-6700HQ​

16​

DF-UD​

256​

320​

96​

96​

22​

4​

1800​

TRUE​

GTX 1650​

4​

i5-9300H​

16​

DF​

128​

256​

64​

64​

22​

6​

780​

TRUE​

GTX 1650​

4​

i5-9300H​

16​

DF​

128​

256​

64​

64​

22​

8​

960​

FALSE​

GTX 1660 Ti​

6​

i7-9700​

16​

DF​

192​

256​

64​

48​

22​

8​

676​

TRUE​

GTX 1070​

8​

i7-7700HQ​

16​

LIAE-UD​

224​

288​

96​

96​

16​

4​

800​

TRUE​

GTX 1070 Ti​

8​

R5-3600​

16​

DF​

192​

256​

64​

64​

22​

10​

1030​

FALSE​

GTX 1080​

8​

i7-8700K​

32​

DF​

192​

256​

64​

64​

22​

8​

780​

TRUE​

GTX 1080 Ti​

11​

W3680​

12​

DF-UD​

288​

256​

80​

80​

20​

12​

1205​

TRUE​

GTX 1080 Ti​

11​

W3680​

12​

DF-UD​

288​

256​

80​

80​

20​

4​

484​

TRUE​

GTX 1080 Ti​

11​

W3680​

12​

DF-UD​

288​

256​

80​

80​

20​

6​

719​

TRUE​

GTX 1080 Ti​

11​

W3680​

12​

DF-UD​

288​

256​

80​

80​

20​

8​

850​

TRUE​

GTX 1080 Ti​

11​

W3680​

12​

DF-UD​

288​

256​

80​

80​

20​

8​

862​

TRUE​

GTX 1080 Ti​

11​

i5-4590​

16​

LIAE​

192​

256​

64​

64​

22​

8​

670​

TRUE​

GTX 1080 Ti​

11​

i5-4590​

16​

LIAE​

192​

256​

64​

64​

22​

12​

900​

TRUE​

Quadro M2200​

4​

E3-1535M v6​

32​

DF​

128​

512​

64​

48​

16​

4​

921​

TRUE​

RTX 2060​

6​

i5-2500K​

8​

DF​

128​

256​

64​

64​

22​

14​

600​

TRUE​

RTX 2060​

6​

i5-2500K​

8​

DF​

160​

256​

64​

64​

22​

6​

2500​

FALSE​

RTX 2060​

6​

R5-2600​

16​

LIAE​

256​

256​

64​

64​

22​

2​

1700​

FALSE​

RTX 2060 S​

8​

R5-3500​

16​

DF-UD​

256​

256​

64​

64​

22​

14​

800​

TRUE​

RTX 2070​

8​

R7-3800X​

16​

DF​

192​

256​

64​

64​

32​

8​

1100​

FALSE​

RTX 2070​

8​

i7-8700​

16​

DF​

144​

256​

64​

64​

22​

8​

400​

TRUE​

RTX 2070 S​

8​

R5-3600​

16​

DF​

192​

256​

64​

64​

22​

5​

600​

FALSE​

RTX 2080​

8​

i7-8700​

16​

DF​

224​

512​

80​

80​

22​

2​

406​

TRUE​

RTX 2080​

8​

i7-8700​

16​

DF​

192​

512​

64​

64​

22​

7​

570​

TRUE​

RTX 2080​

8​

i7-8700​

16​

DF​

192​

512​

80​

80​

26​

3​

570​

TRUE​

RTX 2080​

8​

i7-8700​

16​

DF​

224​

512​

64​

64​

22​

5​

580​

TRUE​

RTX 2080​

8​

R7-3800X​

16​

DF-UD​

320​

256​

64​

64​

22​

5​

478​

TRUE​

RTX 2080 Ti​

11​

i7-9700K​

16​

DF-UD​

256​

256​

64​

64​

22​

20​

800​

TRUE​

RTX 2080 Ti​

11​

i9-9900K​

32​

LIAE-U​

256​

256​

64​

64​

22​

6​

700​

TRUE​

RTX 2080 Ti x2​

22​

R7-2700​

32​

LIAE​

192​

256​

64​

64​

22​

20​

1230​

TRUE​

RTX 3090​

24​

i7-9700K​

16​

DF-UD​

256​

256​

64​

64​

22​

16​

581​

TRUE​

RTX 3090​

24​

R9-3900X​

32​

LIAE-UD​

384​

384​

116​

116​

16​

6​

942​

TRUE​

Tesla P100​

16​

Colab​

16​

DF​

192​

768​

80​

80​

22​

8​

1000​

TRUE​

Tesla P100​

16​

Colab​

16​

DF​

192​

256​

64​

64​

22​

18​

1200​

TRUE​

Tesla P100​

16​

Colab​

16​

DF​

192​

256​

64​

64​

22​

12​

800​

TRUE​

Tesla P100​

16​

Colab​

16​

DF-UD​

256​

320​

96​

96​

22​

4​

460​

TRUE​

Titan RTX​

24​

RT-3970X​

128​

DF​

400​

256​

64​

64​

22​

6​

1350​

TRUE​

Titan RTX​

24​

E5-1650​

64​

DF​

256​

256​

64​

64​

22​

16​

4100​

TRUE​

Titan RTX​

24​

E5-1650​

64​

DF​

224​

256​

64​

64​

22​

20​

4600​

TRUE​

Titan RTX x2​

48​

RT-3970X​

128​

DF​

512​

256​

64​

64​

22​

6​

1700​

TRUE​

Titan RTX x2​

48​

RT-3970X​

128​

DF​

512​

256​

64​

64​

22​

8​

2100​

FALSE​

Titan RTX x2​

48​

RT-3970X​

128​

DF​

400​

256​

64​

64​

22​

12​

2200​

TRUE​

Please use my testing method to measure performance of your configuration, you need to run model twice, once in low and then in high load scenario.
Also make sure you are testing using the latest version of DFL and use the original builds, do not test on forks like MVE one, only iperov's version (unless you disable all additional features).

Low model load (you must test with those values):

RW: enabled
UY: disabled
EMP: disabled
LRD: disabled
GPU Opt on GPU: TRUE
GAN: disabled
Face Style Power: 0 (disabled)
Background Style Power: 0 (disabled)
TrueFace: 0 (disabled, only for DF archis)
Color Transfer: RCT
Clipgrad: FALSE

High model load (you must test with those values):

RW: disabled
UY: enabled
EMP: enabled
LRD: enabled (on GPU)
GPU Opt on GPU: TRUE
GAN: 0.1
GAN Dims: default value (16)
GAN Patch Size: 1/8 of model resolution
Face Style Power: 0.1
Background Style Power: 0 (disabled)
Color Transfer: RCT
Clipgrad: FALSE

If you want to provide additional settings using different paramaters for GAN, GAN DIMS, GAN PATCH SIZE, FSP, BSP, TF, CT, etc you can do so but they must be submitted along with standard testing method results in separate table using the 2nd template as a way to compare both.

Template with example values:

GPU​

VRAM​

CPU​

RAM​

OS​

Page File Size​

Model​

Architecture​

Resolution​

Batch Size (High Load)​

Batch Size (Low Load)​

Iteration Times (High Load)​

Iteration Times (Low Load)​

VRAM Usage Before Training​

VRAM Usage During Training​

AE Dims​

E Dims​

D Dims​

D Mask Dims​

Inter Dims​

Adabelief​

GAN Dims​

GAN Patch Size​

RTX 3090​

24​

i9-13900K​

64​

Windows 11​

256​

SAEHD​

LIAE-UDT​

384​

6​

12​

1000​

500​

1​

23.6​

320​

64​

64​

22​

-​

YES​

16​

48​

Batch Size (typical lowest)​

Iteration Times (Highest Load)​

VRAM Usage During Training​

GPU Opt on GPU​

Adabelief​

LRD​

GAN Dims​

GAN Patch Size​

GAN Power​

Face Style Power​

Background Style Power​

Color Transfer​

True Face​

Clipgrad​

4​

1400​

23.6​

FALSE​

NO​

On CPU​

24​

96​

0.2​

0.001​

0.0001​

LCT​

0.01​

TRUE​

WaveBedo

猛男，你好啊

你有测试过Face Style Power吗？
是不是可以让脸型接近dst，减少虚影啊

Lau9

WaveBedo 发表于 2024-1-20 01:17
猛男，你好啊

你有测试过Face Style Power吗？

是色彩风格更接近dst吧。由于和dst衔接的更平滑，所以边界就弱了。当你羽化的时候体现优势

WaveBedo

Lau9 发表于 2024-1-20 01:25
是色彩风格更接近dst吧。由于和dst衔接的更平滑，所以边界就弱了。当你羽化的时候体现优势 ...

好的

kdq751

WaveBedo 发表于 2024-1-20 01:17
猛男，你好啊

你有测试过Face Style Power吗？

哎，我的3070 淘汰好快啊

oolongpen

4090最高可以跑到什么参数？