Lighthouse XL v2.0

Version 2.0 of the Lighthouse XL LoRA builds upon the successful version 0.9 proof-of-concept by expanding the image set for training, with a larger dataset derived from a curated set of images using AI and my Lora Dataset Curator. This evaluation included over 500 photographic-quality photos of lighthouses. I chose 250 of the top-scoring photos that illustrate the concept of a lighthouse. This improved and optimized training set greatly enhances the ability of any SDXL checkpoint to generate a high-quality image of a lighthouse. I tested the LoRA with several popular SXDL fine-tuned checkpoints with great results (see below).

Notes on Training

I trained the original v0.9 on 100 images, mostly from my personal collection of photographs I have taken over the years. A large majority of those images were taken on black-and-white film, for which the scanning process picked up the film grain. Sometimes, to the detriment of the image. I tried a lot of tricks to reduce the influence of the film grain on the images generated by v0.9, but it appears as banding in some generated images.

Version 2.0 attempts to address the issues in v0.9, beginning with a larger set of training images, specifically replacing the film images that did not scan well. I added images to my collection until I had 500+ quality photographs of lighthouses in diverse environments and lighting conditions. These images were fed to my LoRA Dataset Curator software, which scored each image based on how well it aligns with the concept of "lighthouse." I rejected all images that did not score the maximum value (1.0). This resulted in a set of approximately 375 curated images. From this set, I culled what I deemed to be duplicates based on how they interpreted the concept and settled on a set of 250 training images.

The process produced a better training image set and ultimately, a better LoRA that creates high-quality images under various conditions. Your feedback, whether good or bad, is crucial in determining if my belief is correct.

If you are interested, you can download my LoRA Dataset Curator software and run it on your system. It is entirely free and may help you curate your own concept LoRAs, where the interpretation of the concept is crucial. Using AI to analyze the image helps mitigate some of the biases and subjectivity that can creep in when evaluating images for a particular concept.

Checkpoints Tested

I tested the following checkpoints with the Lighthoue XL v2.0 LoRA <strength:1.0> with the prompt and settings listed below. Each image contains the ComfyUI workflow that generated the image. Drag these images into you ComfyUI environment to reproduce the simple workflow I used.

Prompt: A red and white banded lighthouse sits at the end of a rock wall extending 200 feet from the shore, during sunset. Large waves are crashing against the rocky coast. Seed: 0 CFG: 7.0 Sampler: Euler Scheduler: beta

Checkpoint Without LoRA With LoRA
SDXL 1.0
Juggernaut XL Ragnarok
Hello World XL v7.0
Cyberrealistic XL v7.0
Landscape Realistic Alpha 1.0

Notes on the Test Images

I generated these images on an AMD "Strix Halo" machine with an AMD Ryzen AI Max+ 396 CPU and an integrated Radeon 8060S GPU. Your mileage may vary. In other words, you may never exactly reproduce these images on your system. Welcome to the world of random noise diffusion.

Restrictions and Limitations

I removed the restrictions on generation that I placed on v0.9. I imposed those restrictions simply because there was a possibility that the proof-of-concept was faulty, and I did not want people generating bad images with a faulty LoRA.

I encourage you to create a prompt to generate your lighhouse masterpiece, print it large, and hang it on your wall for everyone to admire.

However, using this LoRA and diffusion models to create images to sell, well that is outside the premissions I give you for use of this LoRA. If you would like to train your own LoRA's and generate images for sale, then by all means go for it. However, using other people's effort and expense to enrich yourself is something you will have to atone for later in life. That goes for all models and LoRA's, not just mine.

Revision History

Version Date Changes CFG Steps Sampler Notes
v2.0 2025-11-17 Re-trained on 250 images 7.0 30 Eular + beta Better quality images, wider context
v0.9 2025-11-04 Initial version, trained on 100 images 7.0 30 Eular + beta Proof of concept