Open JoaoPedroAbubakir opened 11 months ago
What I have tried:
Some other additional info: Nvidia driver version - 546.01
After many hours of discussion with CivitAI discord, we have found a workaround. Using aDetail helps with keeping TI info in metadata and seems to force the loading.
I have this bug aswell. Happens to other embeddings too.
This is most likely a cache problem. If you change something in the promt, the embedding hash will show again, but every consecutive run it will not.
I fixed this locally by disabling the caching in processing.py inside the function get_conds_with_caching(...) by commenting out the lines
for cache in caches:
if cache[0] is not None and cached_params == cache[0]:
return cache[1]
which is probably not optimal, but this also fixes embedding training preview when using text2img promt for me, so it's better than to have this cache since i don't see any negative performance impact or problems with it being disabled.
Is there an existing issue for this?
What happened?
Textual inversions not loading properly. They show up in the first run, then do not for any subsequent run. However, from image result they are not used in either.
Steps to reproduce the problem
What should have happened?
Textual Inversions should have been used to generate the image consistently.
Sysinfo
sysinfo-2023-11-03-07-01.txt
What browsers do you use to access the UI ?
Other: Opera GX
Console logs
Additional information
I have no idea where to log when Embeddings are loaded so I added them to modules.sd_hijack_clip:234, when metadata is generated. You can see from the log that on the first pass it at least showed up, but on the second it dit not.