apologies, my fault. pls disregard my previous tweet, i've picked up the wrong checkpoint. here are 120K 𝑀 samples from @AydaoAI large anime model (aka TANDE) clustered into a set of 256 centroids. 𝘸𝘒𝘡𝘀𝘩 π˜ͺ𝘡 𝘴𝘩π˜ͺ𝘯𝘦 (full collage l4rz.net/aydao-tande-kmeans-…)

4:46 PM Β· Mar 30, 2021

3
4
1
28
Replying to @l4rz @AydaoAI
I am amazed at the quality. Even distorted drawings look intentional and don't appear ugly or uncanny on closer inspection even though they are VERY not-quite-right.
1
0
0
1
And text on pictures even look japanese. Last time I saw something like this, the text immediately betrayed the artificial nature of the pictures, because it was like alphabet from madoka.
1
0
1
2
Replying to @l4rz @AydaoAI
This is blowing my mind. You're sure that every single image in the full-size image you linked is model-generated and not human-created art? 🀯 Some of these images are 🀩
1
0
0
3
it is generated indeed with a lot of effort by a rather determined person (and 5 megawatt-hours or so of compute)
0
0
0
4
Replying to @l4rz @AydaoAI
So did the previous samples had some kind of bias towards the explicitness? did that bias came from underfitting or was it because the model (previous checkpoint) was trained on a different set dataset?
1
0
0
0
tbh since it's not my model, i have no idea what was the composition of training set used for this particular checkpoint. specifically, was the nsfw part of danbooru filtered out or not
1
0
0
2
This Tweet was deleted by the Tweet author.