Noticed that most models produce either non-Indian looking women, or if prompted, produce very dark skin tones, which some, but not all Indian women have. This textual inversion tries to address that. This model produces photo-realistic women with Indian skin-tone, facial features, and body types.
v2 is trained with a larger number of images allowing for more flexibility and is my recommended inversion, unless there's a reason you like the older one.
Suggested settings are in the "About this version".
Produces NSFW images unless used with very low weightages, or forced through prompts like ((clothed)).
A safe for work version of the model can be found here: https://civarchive.com/models/63747/desi-dressed
Description
Updated version of my most popular download. This version of the textual inversion is built with a much larger data set of over 200 pictures. The quality shows.
Like the previous version - this generates NSFW images, unless heavily suppressed through positive and negative prompts.
Recommended settings:
CFG: 7
Steps: 20
Sampler: Euler a
Weight: 0.5-1 (lower weights produce less pronounced features, and hence creates a variety of 'looks').
Comments/ suggestions welcome
FAQ
Comments (13)
This has a Pickle Tensor, don't know if this is safe
That's because it was trained using A1111, which produces only PT files afaik. Will see if I can convert it into a safetensors. Any suggestions on how?
See if this help - https://youtu.be/NNkdhPwrL-g
@Geekybhaijaan Thanks , Have seen this. The issue is all of these are meant for .ckpts. Renaming the .pt to .ckpt and then running the scripts, produces files that are just 16 bytes. So unfortunately, for the moment, .pt is what you're stuck with. There are few scripts that check for safety of the .pt - you're welcome to use those.
@Desi_Cafe Can u suggest those scripts
@Geekybhaijaan Try this: https://github.com/mmaitre314/picklescan
What is the trained keyword for prompt?
For V2: desigirlv2
For V1: desigirl
Where do I save this/How do I use it? Thanks.
Download the file and put it in your 'embeddings' folder of whatever Stable Diffusion repo you are using. Then simply invoke the name of embedding in your positive prompt. Use:
For V2: desigirlv2
For V1: desigirl
You'll see examples in the gallery.
Please note this will only work with Checkpoints that are based on SD1.5. This will not work for SDXL.
how do i use this, i tried pasting the file in embedding folder and using the prompt but it is not working, please help with any resources. or upload safetensor file. thank you
RuntimeError: expected scalar type Half but found Float
what to do with this error? everything else runs fine till i add the trigger word in prompt
No idea, tbh. Did you put this in the embedding folder? Also this file is now more than a year old, so torch etc. may have changed how they use .pt files.