Type | |
Stats | 502 631 |
Reviews | (80) |
Published | Aug 15, 2024 |
Base Model | |
Training | Steps: 3,000 Epochs: 42 |
Usage Tips | Strength: 1 |
Trigger Words | Porsche 911 GTS Spoiler driver numberplate reads XYZ |
Hash | AutoV2 FEEE073A48 |
LoRa of the Porsche 911 GTS 2024.
Generation tipps:
You can generate directly 1920x1080 without problems. It's a bit overtrained but retains more details while inpainting in close ups. Use a lower strenght for initial pass and a higher for detailed cropped (scale 1.5-2x) inpainting for just the car. Use the "spoiler:0-1" trigger if you want to generate with extended spoiler and use "driver" if you want a driver in the car. For custom numberplate text use "numberplate reads <your text here>". Images in the post are all one shot. For Close Ups just put "close up" of "headlights" or "rims".
Some thoughts about training flux:
In my experience (20-30 LoRas...so take it with a grain of salt) flux loves to be trained with higher resolution images (1024) and with a high rank for high detail subjects like cars. Flux like SDXL and SD 1.5 also trains with more coherence if you tell it which perspective of your subject is shown in the images. The quality of the images is more important than anything else. For subject LoRas the ratio of the images doesn't really matter if you want it to be more adaptable but retain close up details train the same images twice once cropped to just your subject and the other one full ratio but with the same caption (no dropout). Keep the background captions concise and only add more detailed captions if there is something out of the ordinary in the background. (ie. lush flowers and trees in a desert).