Type | |
Stats | 1,977 329,512 65.2k |
Reviews | (423) |
Published | Apr 23, 2024 |
Base Model | |
Training | Steps: 736 Epochs: 4 |
Usage Tips | Clip Skip: 1 Strength: 0.5 |
Trigger Words | front_view 2/3_front 1/3_front side_view 1/3_back 2/3_back back_view eye_level higher_angle high_angle +27 more |
Hash | AutoV2 EC2779081C |
What Views & Composition (VnC) does:
Making it easier for adjusting camera angle/viewer angle, character positions (distance to the viewer or the portion within the frame), and the composition of your image.
I made this LoRA because I often struggled with the camera angle and the composition while making images with my favorite SD or SDXL models, especially when using complex prompts.
Usage:
Use different trigger words for different views and/or composition desired.
Explanation to a few trigger words:
1）higher_angle vs high_angle: "higher" means "higher than the eye level", while "high" means "really high" and will usually be closer to the top view (top_down). The same with lower_angle and low_angle.
2）d_short, d_medium, d_long, d_far: referring to the distance of your main subject.
3）comp_left, comp_center, comp_right, etc.: referring to the composition, telling the model that you want the main subject on the left/center/right/top/bottom part of the image.
4）upper_3/4, feet_out: these two actually do the same thing. They are likely to give you a image depicting a person with only their feet out of frame. (Note that "full_body" in this LoRA means real "full" body!)
5）H_symmetrical, L_symmetrical: meaning "horizontally symmetrical" and "longitudinally symmetrical". However, they are NOT meant to produce strictly symmetrical figures, but the approximate symmetry of the layout, so does the "centrosymmetric".
Certain combinations of trigger words could produce better results, eg., "full_body, d_long", or "drone_view, wide_angle".
Although this LoRA is primarily trained for making images of human characters and landscapes, feel free to try it on animals and objects!
I am still exploring the weighs and steps that work best with , so no recommendations at this time.
This is my first LoRA and I have yet a lot to learn. Current (beta) version seems to do better with the views, less responsive to distances, and somewhat confused about L_symmetrical. I will try to improve its performance in the future version(s).
Finally, I hope you find this tool useful!