V2.0 is the recommended version! It's a total overhaul, making generating the "Looking at Self" pose much more consistent and reliable. Thanks to the community feedback re: v1.
Tripled the dataset size with 1,998 training steps for better variety and accuracy.
Intended for 2D, animé or Western illustration style images — supports furry, hetero, and yaoi prompts. Does not work well with realistic images.
(While v2.0 is the better LoRA, v1.0 is still available to allow users to make comparisons, or to preserve nuances in case of user preference.)
This LoRA reinforces the looking at self pose concept for Pony while keeping minimal impact on style.
Enhances the downward head tilt, pulling the chin closer to the chest and revealing more of the crown.
Aligns the subject's eyes with their body when "looking down," avoiding the appearance of "looking away" pensively.
Adjusts pupils toward the nose rather than the center of the eye, for a focused and introspective look.
Tag:
looking at self
Note: the LoRA adds a natural tendency for the subject to touch what they're looking at, such as a body part or clothing.
Trained as a Bounty for user Deletetdyslexi
Description
FAQ
Comments (13)
Doesn't seem to work even when I tested it with AutismMix, and other checkpoints. Going past weight 1.0 causes it to look overfit and it still doesn't get the pose right.
Do you happen to have more examples with the metadata included so I can see if there is something you did differently for testing?
Sorry to hear you're having a problem with it. I tested it with AutismMix without any other LoRAs in play. I've posted more examples with metadata.
I should mention, the training data has many different poses, but the common denominator is that the head and eyes are focused towards the self.
Now sourcing a larger dataset and working on v2.0 to improve consistency.
@RetroSmashtAI That sounds good, I really like the idea of this lora, but even when testing it with the metadata from the results aren't really reproducing the pose. Looking forward to testing v2!~
@LindezaBlue v2.0 is live. Let me know what you think :)
I've been searching for a lora to do this for ages! It's not as reliable as I'd like, even at 1.5 weight it's very prompt-dependant and hit-or-miss... but it's the best we've got and I'm so thankful! All of the diffusion models have way too much preference for the subject to be looking directly "at the camera." It's difficult to overcome! Thank you for this!
Thanks for the feedback. It's true that it doesn't have a 100% success rate. My goal was to reinforce the "looking at self" concept for Pony. After reading users' comments I realise I need to revise this LoRA to bring more consistency to your gens.
@nooooooooooooo v2.0 is live and it's a big improvement. Let me know your thoughts!
Man! This is something I have been searching for ages... even thought about training one myself. I only do realistic images... will give it a shot a post my results...
Other users have mentioned some issues with v1. v2 is in the works
@weilinyangli364 I just uploaded v2.0, it's the recommended version
@RetroSmashtAI Thanks friend! I have ran the model several times with realistic models and no luck... I did not get a single proper results. Now, although I generally do not do anime, I did try an anime model and had better results... maybe the dataset needs to have some realistic images as well for it to work "cross styles"...
@weilinyangli364 Yes, good point. The dataset is exclusively 2D — animé and Western illustrations including furry. It also leans towards NSFW since, most of the time, subjects depicted as looking at themselves are also doing something NSFW. I may add support for realistic images in a future version.
Looks like we don't have an active mirror for this file right now.
CivArchive is a community-maintained index — we catalog mirrors that volunteers upload to HuggingFace, torrents, and other public hosts. Looks like no one has uploaded a copy of this file yet.
Some files do get recovered over time through contributions. If you're looking for this one, feel free to ask in Discord, or help preserve it if you have a copy.