Update README.md
Browse files
README.md
CHANGED
@@ -74,14 +74,20 @@ As same as kxl eps rev2, I add realbooru and pvc figure images for more flexibil
|
|
74 |
|
75 |
## Training
|
76 |
- Hardware: Quad RTX 3090s
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
77 |
- Training
|
78 |
-
- Num Train Images: 8,468,798
|
79 |
-
- Total Epoch: 1
|
80 |
-
- Total Steps: 16548
|
81 |
-
- Training Time: 430 hours (wall time)
|
82 |
- Batch Size: 4
|
83 |
- Grad Accumulation Step: 32
|
84 |
- Equivalent Batch Size: 512
|
|
|
|
|
|
|
85 |
- Mixed Precision: FP16
|
86 |
- Optimizer
|
87 |
- Optimizer: Lion8bit
|
@@ -94,12 +100,6 @@ As same as kxl eps rev2, I add realbooru and pvc figure images for more flexibil
|
|
94 |
- Min SNR Gamma: 5
|
95 |
- Debiased Estimation Loss: Enabled
|
96 |
- IP Noise Gamma: 0.05
|
97 |
-
- Resolution: 1024x1024
|
98 |
-
- Min Bucket Resolution: 256
|
99 |
-
- Max Bucket Resolution: 4096
|
100 |
-
- Other
|
101 |
-
- Caption Tag Dropout: 0.2
|
102 |
-
- Caption Group Dropout: 0.2 (for dropping tag/nl caption entirely)
|
103 |
|
104 |
|
105 |
## Why do you still use SDXL but not any Brand New DiT-Based Models?
|
|
|
74 |
|
75 |
## Training
|
76 |
- Hardware: Quad RTX 3090s
|
77 |
+
- Dataset
|
78 |
+
- Num Images: 8,468,798
|
79 |
+
- Resolution: 1024x1024
|
80 |
+
- Min Bucket Resolution: 256
|
81 |
+
- Max Bucket Resolution: 4096
|
82 |
+
- Caption Tag Dropout: 0.2
|
83 |
+
- Caption Group Dropout: 0.2 (for dropping tag/nl caption entirely)
|
84 |
- Training
|
|
|
|
|
|
|
|
|
85 |
- Batch Size: 4
|
86 |
- Grad Accumulation Step: 32
|
87 |
- Equivalent Batch Size: 512
|
88 |
+
- Total Epoch: 1
|
89 |
+
- Total Steps: 16548
|
90 |
+
- Training Time: 430 hours (wall time)
|
91 |
- Mixed Precision: FP16
|
92 |
- Optimizer
|
93 |
- Optimizer: Lion8bit
|
|
|
100 |
- Min SNR Gamma: 5
|
101 |
- Debiased Estimation Loss: Enabled
|
102 |
- IP Noise Gamma: 0.05
|
|
|
|
|
|
|
|
|
|
|
|
|
103 |
|
104 |
|
105 |
## Why do you still use SDXL but not any Brand New DiT-Based Models?
|