| 30 Nov 2024 |
|  @jambyowl3241:unredacted.org joined the room. | 07:39:54 |
|  0xfn joined the room. | 09:25:42 |
 primere | 
Download Image_A685_20243011_103700_705162941572791_1632x2560_1.jpg | 11:24:56 |
| primere | 
Download Image_A663_20243011_112309_339367444131064_2560x1632_1.jpg | 11:24:56 |
| primere | 
Download Image_A622_20243011_113254_454292930779151_1632x2560_1.jpg | 11:24:57 |
| primere | 
Download Image_A666_20243011_115115_719667533254051_1632x2560_1.jpg | 11:24:57 |
| primere | 
Download Image_A689_20243011_120806_55620384133763_1632x2560_1.jpg | 11:24:58 |
| primere | 
Download Image_A696_20243011_121058_674127531614932_2560x1632_1.jpg | 11:24:58 |
| primere | 
Download Image_A690_20243011_110658_970727927776657_1632x2560_1.jpg | 11:24:59 |
| primere | 
Download Image_A654_20243011_120508_863252474856079_1632x2560_1.jpg | 11:24:59 |
| primere | Pixart sigma several sci-fi and architecture prompts: | 11:24:59 |
|  @uuuuuuuuuuhbg:matrix.org joined the room. | 11:38:45 |
| @uuuuuuuuuuhbg:matrix.org left the room. | 11:55:35 |
|  furkangozukara changed their profile picture. | 14:32:16 |
|  @a:noevil.pl changed their display name from 🟡1️⃣3️⃣🟢 to 🟡1️⃣3️⃣🟢 #️⃣🇷🇹🇲🇲🇩🇦🇾🇸. | 16:11:18 |
| @a:noevil.pl changed their display name from 🟡1️⃣3️⃣🟢 #️⃣🇷🇹🇲🇲🇩🇦🇾🇸 to 🟡1️⃣3️⃣🟢 🖕#️⃣🇷🇹🇲🇲🇩🇦🇾🇸🖕. | 16:12:25 |
| @a:noevil.pl changed their display name from 🟡1️⃣3️⃣🟢 🖕#️⃣🇷🇹🇲🇲🇩🇦🇾🇸🖕 to 🟡1️⃣3️⃣🟢 🖕😠🖕 #️⃣🇷🇹🇲🇲🇩🇦🇾🇸. | 16:16:58 |
| @a:noevil.pl | In reply to @japtain.cack:matrix.talos.mimir-tech.org
Do you have any pointers for dialing in PAG? What am I looking for visually to determine an increase/decrease in PAG? PAG is effectively an extra cfg done over positive prompt and increasing scale equals to increasing positive prompt strength | 16:50:50 |
| @a:noevil.pl | Regular cfg performed like negative + (positive - negative) * cfg | 16:51:05 |
| @a:noevil.pl | * Regular cfg performed like pred = negative + (positive - negative) \* cfg | 16:52:44 |
| @a:noevil.pl | * Regular cfg performed like pred = negative + (positive - negative) * cfg | 16:52:47 |
| @a:noevil.pl | PAG is pred + (positive - pag) * scale | 16:53:25 |
| @a:noevil.pl | * PAG is pred = pred + (positive - pag) * scale | 16:53:39 |
| @a:noevil.pl | Where pag is the positive modeled with middle block disabled | 16:54:19 |
| @a:noevil.pl | It does not mean final cfg equals cfg + scale but you will notice image getting more saturated as you increase scale | 16:55:11 |
| @a:noevil.pl | But decreasing cfg might actually let you go for higher scale values | 16:55:58 |
| @a:noevil.pl | I should read paper about why doing that extra guidance with skipped blocks improves quality | 16:56:35 |
| @a:noevil.pl | Recent changes allow you to perform the same trick over DiT models (including stable audio) | 16:57:05 |
| @a:noevil.pl | In reply to @a:noevil.pl
It does not mean final cfg equals cfg + scale but you will notice image getting more saturated as you increase scale Oh and if your cfg is low enough you will also notice negative prompt becoming weaker obviously | 16:58:01 |
 | it goes well with SEG | 17:20:30 |
