Type | |
Stats | 93 74 |
Reviews | (7) |
Published | Feb 3, 2024 |
Base Model | |
Training | Steps: 12,600 Epochs: 6 |
Usage Tips | Strength: 0.6 |
Trigger Words | frdbrnc SUV |
Hash | AutoV2 6539934820 |
This LoRA is meant to generate more accurate images of the most recent Ford Bronco. It's my first LoRA and first Civitai submission so feedback is definitely welcome! Here's how it was trained:
Used kohya on a Windows machine with a 10GB 3080
Class: SUV
Instance: frdbrnc
36 sources images mostly resized to 1152x768 using birme.net
Used GIT captioning with hand-edits to add things like "3/4 view", "side view", "front view", "hard top", "soft top", "top down" and also added "ford bronco" to each caption file.
Trained on Realistic Vision V6.0 B1 - V5.1 (VAE) | Stable Diffusion Checkpoint | Civitai
60 repeats, 6 Epochs
Tips:
I get great results with DPM++ SDE Karras, 4-6 CFG, 30-70 steps
0.6 LoRA strength
768x512 or 768x768 helps ensure you don't get a cropped vehicle
Seems to play nicely with other LoRAs
RealisticVision, revAnimated, and CyberRealistic seem to like it
A simple "ford bronco" produces good results but experiment with adding the instance/class strings of 'frdbrnc' and/or 'SUV' as well
Other Notes:
Only problem seems to be a halfway interpretation of a grille guard bar. Just experiment with different step counts and it should go away without too much trouble.