DFL 2.0 Model Settings and Performance Sharing
Old spreadhseet (won't be updated unless users start to share settings more often):
In this thread you can share and lookup performance of models at specific settings on various hardware configurations.
Please use my testing method to mesasure performance of your configuration, you need to run model twice - once in low and then in high load scenario.
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
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: 16
GAN Patch Size: 1/8 of model resolution
Face Style Power: 0.1
Background Style Power: 0 (disabled)
TrueFace: 0.01 (only for DF archis)
Color Transfer: RCT
If you want to provide additional settings using different paramaters for GAN, GAN DIMS, GAN PATCH SIZE, FSP, BSP, TF and CT you can do so but they must be submited along with standard testing method results.
Template:
GPU and VRAM: XX
CPU: XX
RAM: XX
OS: XX
Model Type, Architecture, Face Type: XX XX XX
Resolution: VALUE
Batch Size: low load - high load
Models_opt_on_gpu: TRUE/FALSE
Iteration time: low ms - high ms
AE_Dims: VALUE
E_Dims: VALUE
D_Dims: VALUE
D_Mask_Dims: VALUE
LR_Dropout: TRUE/FALSE/CPU
AdaBelief Optimizer: TRUE/FALSE
True_Face: YES/NO or VALUE
GAN (power): 0.1
GAN Dims: 16
GAN Patch Size: 1/8
Face_style: YES/NO or VALUE
Bg_style: YES/NO or VALUE
CT_Mode: NONE/TYPE
Clipgrad: TRUE/FALSE
NOTES: XX
DFL Commit/Version: XX
Example of model setting submission using the provided template:
Old spreadhseet (won't be updated unless users start to share settings more often):
In this thread you can share and lookup performance of models at specific settings on various hardware configurations.
Please use my testing method to mesasure performance of your configuration, you need to run model twice - once in low and then in high load scenario.
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
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: 16
GAN Patch Size: 1/8 of model resolution
Face Style Power: 0.1
Background Style Power: 0 (disabled)
TrueFace: 0.01 (only for DF archis)
Color Transfer: RCT
If you want to provide additional settings using different paramaters for GAN, GAN DIMS, GAN PATCH SIZE, FSP, BSP, TF and CT you can do so but they must be submited along with standard testing method results.
Template:
GPU and VRAM: XX
CPU: XX
RAM: XX
OS: XX
Model Type, Architecture, Face Type: XX XX XX
Resolution: VALUE
Batch Size: low load - high load
Models_opt_on_gpu: TRUE/FALSE
Iteration time: low ms - high ms
AE_Dims: VALUE
E_Dims: VALUE
D_Dims: VALUE
D_Mask_Dims: VALUE
LR_Dropout: TRUE/FALSE/CPU
AdaBelief Optimizer: TRUE/FALSE
True_Face: YES/NO or VALUE
GAN (power): 0.1
GAN Dims: 16
GAN Patch Size: 1/8
Face_style: YES/NO or VALUE
Bg_style: YES/NO or VALUE
CT_Mode: NONE/TYPE
Clipgrad: TRUE/FALSE
NOTES: XX
DFL Commit/Version: XX
Example of model setting submission using the provided template:
Code:
GPU and VRAM: RTX 3090 24GB
CPU: Core i9 11900k
RAM: 64GB
OS: Win11
Model Type, Architecture, Face Type: SAEHD LIAE-UDT WF
Resolution: 448
Batch Size: Low 10 - High 6
Models_opt_on_gpu: TRUE
Iteration times: Low 850ms - High 1400ms
AE_Dims: 416
E_Dims: 112
D_Dims: 112
D_Mask_Dims: 22
LR_Dropout: TRUE
AdaBelief Optimizer:: TRUE
True_Face: NO
GAN:: 0.1
GAN Dims: 56
GAN Patch Size: 16
Face_style YES/NO: 0.1
Bg_style YES/NO: 0.1
CT_Mode NONE/TYPE: RCT
Clipgrad TRUE/FALSE: FALSE
NOTES: "It just works" - Todd Howard
DFL Commit/Version: January 1st 2022
Last edited:
