Selfie time all the time.
V2 Update: Re-trained data as a Lycoris, trained in new trigger words.
Realism models tend to resist a selfie pov sans camera, even with Controlnet. This Lora aids poses where one or both arms appear to be reaching behind the camera, with the camera/phone out of frame. Works with one person or couples. Also helps OpenPose controlnet allow hands to be out of frame.
Try with 0.7-0.85 weight, "armoutstretched" trigger word, more emphasis may be required for more complex prompts. Also try adding "pov view" to assist.
Description
Trained on 165 images of people taking selfies of themselves without a visible camera, aka reaching out of frame. May be slightly undertrained. All examples prompted using Realistic Vision v1.3, then upscaled.
Training notes:
Trained using Dreambooth extension in WebUI 2.1.6.
DIM: 64, TR: 32
LR: 0.00008, Text LR: 0.000008, polynomial
Bucketing, max res. 1024
Uses [filewords]
No regularization
FAQ
Comments (5)
Great idea! Very creative :)
Thanks!
Thank you!
facetime, messaging, app all seem to help a little.
as do close up, personal, intimate etc.
"self-portrait photography" is also a thing (for bringing the wrist close instead of just hiding the hand out of frame no matter how far it actually is from the camera.) give the sampler every excuse it wants to
POV seems to help even more when it's one of those "POV you are...." fragments you see a lot these days. also consider that "8K Canon DSLR"s besides being very heavy, do not produce the type of image associated with selfies. even at a very basic level - phones are noisy while ILCs have shallow DOF. high frequency (angular resolution) vs low. more fundamental than electricity to signal processing. Nyquist never died, he lives on in your safetensors.
just for the love of god don't say the Ph- word because once there's one in the scene you can never get rid of it
there's probably an embedding waiting to be discovered here. certainly there's one which describes the distinction between selfies and photos of people taking selfies - good luck explaining mirrors to a machine vision/image generation model :)
Excellent extra prompt words ideas!
I re-trained the model and added some of your suggestions as trigger words. I think it helped!