This is my first public LoRa model is designed to create an extremely low angle view from below, with the feet closer to the viewer and the rest further away.
I trained it on iLLustrious v01. I tested it on it and on WAIillustriousSDXL v16. It may not work on other checkpoints (especially realistic ones on a different database).
45 images were used at low and very low angles. Some were rotated 90, 180, and 270 degrees (probably not the best idea).
English isn't my native language, so I'm using a translator. I hope all my explanations have been translated clearly (I'll double-check, but you never know).
Warning: Characters may have excessively large thighs and large shins. Legs may be crooked. Characters may attempt to dress themselves if clothing isn't specified. And much more.
Add in negative promt: fisheye, lens effect, (multiple legs:1.1) - Perhaps this will help to eliminate some of the problems.
My prompts don't all work as expected, but I'll list them:
Main trigger: "ultra0Below2View"
Important prompt #1: "from below, low angle view" - without these, it often renders without tilting or zooming in on the bottom.
Important prompts #2: "towering figure, foreshortening" - If the angle is very low. Although, judging by the tests, the word "figure" was unnecessary.
Camera position: "front view, back view, side view" - for the front view (facing you), the view where the backside is visible but not the face, and the side view. It works inconsistently.
Camera position (IMPORTANT): "thighs in foreground, knees in foreground, shins in foreground, ankles in foreground, feet in foreground". Sometimes it doesn't position the camera if, for example, you specify shins, but the prompt contains shoes or heels. Сhoose one thing.
Body direction: "turn1north" - up, north; "turn1south" - down, south; "turn1west" - right, west; turn1east - east, right. This is for directing the entire character in the desired direction. It doesn't seem to work at all due to the peculiarities of SDXL (even QWEN gets confused when analyzing the art or explaining it). However, sometimes it works, with high values of 0.8-1.
Image rotation: "rotated view". Well, it's related to turn1north above. And it's also unstable, or maybe even non-functional.
P.S. For those who are having problems with my model, but need something similar right now, I found a couple of models (thanks to them for their work):
1) ground upperspective (Illustrious) https://civarchive.com/models/2641210/ground-upperspective?modelVersionId=2965640
2) From Below/Low Pov (Illustrious) https://civarchive.com/models/1304809/from-belowlow-pov-illustrious
3) Foreshortening Slider LoRA | Illustrious XL https://civarchive.com/models/1918778/foreshortening-slider-lora-or-illustrious-xl
4) Wide Power Stance Low-Angle Shot | POSE https://civarchive.com/models/2043416/wide-power-stance-low-angle-shot-or-pose
Description
Updated version. 1.2 right away, as 1.1 had a minor issue. Overall, I used 57 images for training, step 30, clip 2.
Now it works a little better on other checkpoints, but there are noticeable issues with realistic ones (which have nothing to do with illustrious 0.1). Use img2img for the ultra-low angle under the feet—I couldn't get it to work with txt2txt.
Tips: I've noticed that a variety of words can affect the result (depending on the checkpoint), so start with a small prompt and a minimal negative prompt.
And add in negative promt: fisheye, lens effect, (multiple legs:1.1)
The tags remain the same, I just adjusted them and distributed them better (explanations in version 1.0):
Key tag: ultra0Below2View;
Important tags: from below, low angle view, foreshortening;
Direction: front view, back view, side view;
Distance: thighs in foreground, knees in foreground, shins in foreground, ankles in foreground, feets in foreground;
Direction (seems not to work): turn1north, turn1west, turn1south, turn1east, rotated view





