high weighting (>=4) causes some really messed up images #6325
Replies: 4 comments 3 replies
-
Why you use such large weight? What are you trying to achieve? |
Beta Was this translation helpful? Give feedback.
-
My suspicion is that depending on the (unknown internal) weight of the prompt you use, in this case samdoesarts, you can find and exceed a threshold where it will blow away almost all other prompt weights, leaving you with results like you found. I think you found it somewhere in the 2.5-3.5 range for your example, as even that 3.0 is starting to look like your prompt is overconverged. I've actually found a few embeddings that look fine at (1.0), but with even (1.1) they completely overpower and distort the entire image. One of those "you never really know what's under the hood until you try" quirks of AI. Lot of trial an error. |
Beta Was this translation helpful? Give feedback.
-
I have a similar issue. since an update a while ago weights over 2.0 are iffy. |
Beta Was this translation helpful? Give feedback.
-
the weight is relative to the rest of your prompt. So if you have 10 words in your prompt and have one word with a higher weight, it is more significant weight increase compared to if you had 100 words with one word of the same higher weight. So for your 200 weight, (which is crazy crazy crazy high) did you have a ridiculous size of the prompt? |
Beta Was this translation helpful? Give feedback.
-
I have a prompt, and I used
(samdoesarts:3)
which resulted in:then updated it to
(samdoesarts:4)
which resulted in (same seed):Same seed. This is weird. Not sure if this is a bug, known issue, user error, etc. It was surprising though.
I'm running SD 2.1 and Automatic1111
3e22e294135ed0327ce9d9738655ff03c53df3c0
(updated last night)Beta Was this translation helpful? Give feedback.
All reactions