embedding pt file doesnt work

[1blackbar]

I finetuned the face of stallone but my pt file just does not get detected in webui like my other pt files, this one is bigger at 153kb, others that work are 6kb, it was finetuned at 256 res and samples looked very promising
Any way to get it working ? Yeh its actual output from SD , not a real photo

Here is the file
embeddings_gs-14499.zip

[AUTOMATIC1111]

this is a [50, 768] tensor. How is this supposed to work? Do you insert all 50 into the prompt? What code did you train with to get 50?

[1blackbar]

i trained with 50 vectors enabled, so this wont work ?whats the maximum i can get away with ? I can start ovet just need to know whats the max vector amount i can do
this repo https://github.com/nicolai256/Stable-textual-inversion_win

target: ldm.data.personalized.PersonalizedBase
params:
size: 256
set: train

personalization_config:
  target: ldm.modules.embedding_manager.EmbeddingManager
  params:
    placeholder_strings: ["*"]
    initializer_words: ["face","photo"]
    per_image_tokens: false
    num_vectors_per_token: 50
    progressive_words: False

[AUTOMATIC1111]

no no don't I'm trying to figure it out.

[1blackbar]

Great, cause results are pretty incredible on some samples , he looks very bad in plain SD weights , these are i think 7 epochs and 11k iters, it worked and gave me samples during training so it should work somehow with SD , right ?
So 50 vectors is like "taking over" 50 words from a prompt ? SO it will be very hard to edit a style of it
They say that only DDIM sampler works with it ? here
rinongal/textual_inversion#35

[AUTOMATIC1111]

oh man you are going to be so disappointed with this

[AUTOMATIC1111]

[AUTOMATIC1111]

I'm going to try and test the embedding using lstein's repo to make sure those are the right results

[1blackbar]

Its not that bad, still a test, are they even edtable ? i dont think so, will you update the repo so i can try it? also can i do with 20 vectors ? Heres more, maybe previous ones were better, i stopped when noticed degradation, i planned to resume it with 512 which should get rid of artifacts , i think during outputting samples its going by 100 steps or even 200 ddim sampler
Reults during training did not had these artifacts tho
checkpoints.zip

[AUTOMATIC1111]

It's up, experiment away. I did a stupid thing putting -1 as a placeholder to a place where I shouldn't have, and spent more than an hour trying to catch the phantom error that this cause.

[1blackbar]

Hey happened to me when coding a lot of times, were just human , haha
Tsted it, works great, i got some good results with higher steps about 100 hes like on my examples, also hse pretty much uneditable from photo style, but you can make him smile ,yelll etc, i need to figure out a balance between editability and identity preservation. I can put him into other movies tho but its hard