This is trained with erasing https://github.com/rohitgandikota/erasing with negative negative_guidance and negative start_guidance. This pulls the model towards concepts. It has no dataset, it was only self trained on negative erasure.
Training keywords:
big head mode, full body shot, gigantic head, small body
This is based on https://civarchive.com/models/13565/criarcys-fantasy-to-experience . Marked nsfw because the base is nsfw
Proof of concept. Will not receive updates.
Description
FAQ
Comments (8)
This would work well as a Lora
This is fantastic. [See comments for details on how to run!]
This technique could be a major opportunity to refine other models, so your experience is valuable.
https://github.com/ntc-ai/conceptmod/
I am putting it together there.
@ntc Awesome, thank you! This answers all of my questions.
I'd gotten the environment running, but was indeed wondering about various argument values (number of iterations, start_guidance/negative_guidance == -1 or a different negative value, etc).
@zurichcreator I'm not sure the optimal values, it just now started working and the test cycle is hours. I just released two new models trained with conceptmod and the params specified in the `train_sequential.sh`
Glad you got it running!
@ntc I trained a model...but can't for the life of me get it from .pt to .ckpt format, even after trying various scripts on the web. I see the instructions to directly generate images from the CLI but would love to use some of my other workflows.
Would you mind adding to your GitHub repo README, or here, how you made that final conversion back to a .ckpt? Easy to go from there to safetensors, of course.
@zurichcreator pt and ckpt are the same so just rename it I think
That totally worked. Thanks so much for your patience and time - I've been really excited about your concept here and it feels great to have overcome the issues with your help. Really appreciate your contributions to this community already, and looking forward to any future models you post!
One feature idea: it would be quite helpful if intermediate images are saved out as the model is training, which could help guide tuning of the guidance parameters and also assessing whether a particular prompt is working as intended.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.