Type | |
Stats | 970 51,181 |
Reviews | (89) |
Published | Nov 12, 2023 |
Base Model | |
Trigger Words | antiSketchV2 |
Hash | AutoV2 DF77382623 |
Just like antiBlur pushes image away from not completed and low detailed artwork that got into training data of the SD. Impacts shading, small details, background quality and fingers count. Preview images were not post-processed (e.g. img2img or inpaint).
My next goal is to work on antiBlur.
=====================================================================
V1
Preview images use V_0_2_2 which eventually became V1 - it came through several iterations before being released.
V1_1
Just a small tweak, but it resulted in positive impact on small details quality and slightly more correct anatomy and more pronounced eyes.
V2
This version jump was quite time and resource consuming. There was 8 iterations before I and my gf agreed that this is better than the previous (in average, though, not overwhelmingly).
Technically, there were just two more vectors added. "Just two!? And that took you a couple months?" - one may ask. Well, that it because the idea of this embedding is to be more like a scalpel, not a hammer. The hardest part was to tweak weights to achieve three things: be better than the previous, have the least impact on the composition and be useful for each and every occasion.
Keep in mind, that V2 can noticeably change composition. The third comparison was chosen specifically to show this. In other words, it MAY be so that if you try to recreate some image made with V1_1, the result would be different.