update filters4 filter with the new trained multi-lr&augment filters

mnoiy · xiaoxial · commit d42e9f4e78c8 · 2023-11-22T03:21:41.000+08:00
diff --git a/filters4/Qfactor_cohbin_2_10 b/filters4/Qfactor_cohbin_2_10
@@ -1,2 +1,2 @@
-0.227554
-0.473945
+0.194645
+0.425714
diff --git a/filters4/Qfactor_cohbin_2_10_2 b/filters4/Qfactor_cohbin_2_10_2
@@ -1,2 +1,2 @@
-0.155129
-0.335069
+0.142225
+0.346073
diff --git a/filters4/Qfactor_cohbin_2_8 b/filters4/Qfactor_cohbin_2_8
@@ -1,2 +1,2 @@
-0.215022
-0.446761
+0.192916
+0.405942
diff --git a/filters4/Qfactor_strbin_2_10 b/filters4/Qfactor_strbin_2_10
@@ -1,2 +1,2 @@
-0.000184
-0.001952
+0.000270
+0.002110
diff --git a/filters4/Qfactor_strbin_2_10_2 b/filters4/Qfactor_strbin_2_10_2
@@ -1,2 +1,2 @@
-0.000845
-0.004634
+0.000569
+0.003027
diff --git a/filters4/Qfactor_strbin_2_8 b/filters4/Qfactor_strbin_2_8
@@ -1,2 +1,2 @@
-0.001044
-0.018442
+0.001269
+0.022169
diff --git a/filters4/config b/filters4/config
@@ -1 +1 @@
-24 3 3 11
+24 3 3 11
diff --git a/filters4/filterbin_2_10 b/filters4/filterbin_2_10
diff --git a/filters4/filterbin_2_10_2 b/filters4/filterbin_2_10_2
diff --git a/filters4/filterbin_2_8 b/filters4/filterbin_2_8
diff --git a/filters4/filternotes.txt b/filters4/filternotes.txt
@@ -1,4 +1,7 @@
-1. Filters trained on 675 images of pexels with bilinear degradation
+1. Filters trained on 675*4 images of pexels with area & bicubic & lancoz & nearest degradation
 2. Angle=24 Strength=3 Coherence=3 setting for both filters
-3. Cheapupscale: using output with bilinear upscaling for 1st pass, using output from 1st pass filter as cheapupscale of 2nd pass.
-4. 1st and 2nd pass filters both trained on higher sharpened HR as reference (s=1.5)
+3. Cheapupscale: using output with bilinear upscaling for 1st pass, No cheapupscale of 2nd pass training.
+4.  10bit 1pass filter trained with multi-lr: bicubic+area+nearest+lanczos as LR and HRSharpen50 as HR
+    10bit 2pass filter trained with  HR after single augment:blur as cheapupscale images and sharpened HR images as HR
+    8bit 1pass filter trained with multi-lr: bicubic+area+nearest+lanczos as LR and HRSharpen50 as HR
+    8bit 2pass filter is the same with the filter4-2pass filter of the past version.
