Transformers
English
controlnet
Gerold Meisinger commited on
Commit
45cc134
·
1 Parent(s): 589d18e

control-edgedrawing-cv480edpf-rect-fp16

Browse files
README.md CHANGED
@@ -142,17 +142,19 @@ see experiment 3. cleaned original images following the [fastdup introduction](h
142
  321 too bright
143
  57 blurry
144
  68621 unique removed (that's 38%!)
 
 
145
  ```
146
 
147
- restarted from 0 with left-right flipped images and `--mixed-precision="no"` to create a master release and converted to fp16 afterwards.
148
 
149
- **Experiment 6.0 - control-edgedrawing-cv480edpf-rect-fp16-checkpoint-XXXXX**
150
 
151
  see experiment 5.0.
152
- * included images with aspect ratio > 2
153
  * resized images with shortside to 512 which gives us rectangular images instead of 512x512 squares
154
- * center-cropped images to 512x(n)*64 (to make them SD compatible) and max longside 1024
155
- * sorted duplicates by `similarity` value from `laion2b-en-aesthetics65` to get the best `text` from all the duplicates
 
156
 
157
  ```
158
  183410 images in total
@@ -162,12 +164,25 @@ see experiment 5.0.
162
  436 too bright
163
  31 blurry
164
  76288 unique removed (that's 42%!)
 
 
165
  ```
166
 
 
 
167
  restarted from 0 and `--mixed-precision="fp16"`.
168
 
 
 
 
 
 
 
 
 
169
  # Ideas
170
 
 
171
  * make conceptual captions for laion
172
  * integrate edcolor
173
  * try to fine-tune from canny
 
142
  321 too bright
143
  57 blurry
144
  68621 unique removed (that's 38%!)
145
+ ------
146
+ 111589 unique images (x2 left-right flip)
147
  ```
148
 
149
+ restarted from 0 with left-right flipped images and `--mixed-precision="no"` to create a master release and convert to fp16 afterwards.
150
 
151
+ **Experiment 6.0 - 2023-10-02 - control-edgedrawing-cv480edpf-rect-fp16-checkpoint-45000|90000|135000**
152
 
153
  see experiment 5.0.
 
154
  * resized images with shortside to 512 which gives us rectangular images instead of 512x512 squares
155
+ * included images with aspect ratio > 2
156
+ * center-cropped images to 512x(n)*64 | n=8..16 , which keeps them SD compatible
157
+ * sorted duplicates by `similarity` value from `laion2b-en-aesthetics65` to get the "best" `text` from all the duplicates according to laion
158
 
159
  ```
160
  183410 images in total
 
164
  436 too bright
165
  31 blurry
166
  76288 unique removed (that's 42%!)
167
+ ------
168
+ 107122 unique images (x2 left-right flip)
169
  ```
170
 
171
+ 1 epoch = 107122 * 2 / 4 = 53561 steps per epoch
172
+
173
  restarted from 0 and `--mixed-precision="fp16"`.
174
 
175
+ TODO: Why did I end up with less images after I added more images? fastdup suddenly finds even more duplicates. Is fastdup default threshold=0.9 too aggressive?
176
+
177
+ **Experiment 6.1 - control-edgedrawing-cv480edpf-rect-fp16-gas16-checkpoint-XXXXX**
178
+
179
+ see experiment 6.0. `--gradient_accumulation_steps=16`.
180
+
181
+ 1 epoch = 107122 * 2 / 16 = 13390 steps per epoch
182
+
183
  # Ideas
184
 
185
+ * experiment with higher gradient accumulation steps
186
  * make conceptual captions for laion
187
  * integrate edcolor
188
  * try to fine-tune from canny
control-edgedrawing-cv480edpf-rect-fp16-checkpoint-1350000.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3700c770639bf383294cd58e94672ba4e4d8a9e1a40dfa13d2b60bceb345fef3
3
+ size 722598616
control-edgedrawing-cv480edpf-rect-fp16-checkpoint-45000.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3baceacb0c118266c9ae301820bf65cce0127a7189f8ddc90f0d64fd5f43f981
3
+ size 722598616
control-edgedrawing-cv480edpf-rect-fp16-checkpoint-90000.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2263d0f34a55ac9e73eb4ebefff5d7faadb235447fa1586147d18281796c5e1f
3
+ size 722598616
control-edgedrawing-cv480edpf-rect-fp16-checkpoint-all.zip ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d94593a8df38ca945b97388c3bb344edc28099490511910410078dfe18b75491
3
+ size 222528478