stable-diffusion-webui icon indicating copy to clipboard operation
stable-diffusion-webui copied to clipboard

Embeddings TI are not working - not used by prompt

Open Centurion-Rome opened this issue 2 years ago • 4 comments

Describe the bug A clear and concise description of what the bug is.

To Reproduce Steps to reproduce the behavior:

  1. put trained embedding in "embedding" folder or train new one and start web UI
  2. Go to txt2image and enthe ter prompt that uses your embedding (person in my case)
  3. Click on generate
  4. See that your embedding was not used when generating the image Expected behavior Generated image used embedding from prompt

Screenshots

Desktop (please complete the following information):

  • OS: Windows
  • Browser chrome
  • Commit revision 1b91cbbc1163d3613aa329bed3aecd8e29ca52ca

Additional context

Centurion-Rome avatar Oct 19 '22 07:10 Centurion-Rome

Do you use the filename of the embedding in your prompt? Is the embedding ( a .pt or a .bin file) placed in the embeddings folder? Because it is working for me, chrome, windows.

buzzz3d avatar Oct 19 '22 09:10 buzzz3d

What's the name of your embedding? I think it should be relatively unique, otherwise you'll get weird results, since some training data included your non-unique name and it results in overlapping of concepts. It worked for me like 5-10% of the time when i used name.pt and in other cases it generated empty buildings or some patterns, but when i renamed it to name123.pt it became a lot more stable. But i'm no expert in those, dont take my word for it.

mezotaken avatar Oct 19 '22 10:10 mezotaken

What's the name of your embedding? I think it should be relatively unique, otherwise you'll get weird results, since some training data included your non-unique name and it results in overlapping of concepts. It worked for me like 5-10% of the time when i used name.pt and in other cases it generated empty buildings or some patterns, but when i renamed it to _name_123.pt it became a lot more stable. But i'm no expert in those, don't take my word for it.

Interesting thing. My embedding has name "name123_ti.pt" (5 vectors) So prompt "a photo of name123_ti" generated a random picture without embedding

prompts like "a photo of name123_ti as hero" or "a photo of name123_ti as marble statue" uses embedding.

Centurion-Rome avatar Oct 19 '22 17:10 Centurion-Rome

Some further investigations:

  1. Euler a, CFG Scale 7,5 and 15 or 20, Steps 35, with prompt "photo of name123_ti" generated a random picture without embedding
  2. Euler a, CFG Scale 21,Steps 35 with prompt "photo of name123_ti" generated picture with embedding (sometimes)

Embedding trained 65000 steps with 5 vectors - prveious from training worked as expected

Centurion-Rome avatar Oct 19 '22 17:10 Centurion-Rome

Closing as stale.

catboxanon avatar Aug 03 '23 16:08 catboxanon