The camera angles in the training dataset are oriented either perfectly aligned with gravity or pointing directly in the opposite direction.
I captioned the perfectly top-down images as bird's eye view and the perfectly bottom-up images as worm's eye view, then trained the model on anima_baseV10.
Tip for FT Models: When using this LoRA with already fine-tuned (FT) models, you can achieve the same intense effect by heavily boosting the weights of the standard tags, such as using (from_above:3) or (from_below:3).
Description
Initial.
Known Issues: When using the trigger words bird's eye view or worm's eye view, the model may accidentally generate literal "eyes" or "worm-like monsters" in the image due to word associations.
To bypass this, please use from above / from below instead, or use the negative prompt to filter out and remove those unwanted artifacts.
Used settings:
dim8/alpha8, unet 1e-05(adamw8bit/cosine_with_restarts[2]), 100images / 1000 steps
Extra network args:
block_dims=0,0,0,0,4,4,8,8,8,8,8,4,4,0,0,0,0 block_alphas=8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8 conv_dim=0 conv_alpha=1 rank_dropout=0.25 module_dropout=0.1
FAQ
Comments (1)
This is awesome, Anima really struggles with perspective so I hope you keep making loras like this!
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.