Type | |
Stats | 312 46 270 |
Reviews | (46) |
Published | Sep 17, 2024 |
Base Model | |
Training | Steps: 3,000 |
Trigger Words | XRAY x-ray |
Hash | AutoV2 D4F0FAE40C |
The model is trained on high resolution x-ray images of humans, dogs, other animals, insects, fish, flowers and electronic parts like laptops or mobile phones. Opening a variety of prompts.
Mildly important is the fact that the training data was originally in the 6-18K resolution range. I downscaled it using Lanzos in 3-4K and applied minor post sharpening because of the nature of down scaling while keeping the sampling progressive.
I trained various versions using the same data set, but the prodigy optimizer worked the best for this particular style/topic.
I had some test renders where a colored photo look bleeds through, but generally what was surprising is how well it often retains a distinct 2D look of an x-ray image considering how much Flux loves DOF.