Realistic game LoRa-XL edition prop icons - Realistic v1.0

This time I used A800-80G VRAM with 637 images as source material, using the XL default checkpoint, training for 60 epochs, saving an XL-lora model every 6 epochs, with other parameters set to default, totaling 403,800 steps.

The actual results are really good because the upgraded optimization of XL has significantly improved the text recognition ability. The generated text content can basically meet the accuracy requirements (although there are random variations and styling issues).

It not only learns the effects of existing materials in the training set, but also can transfer styles and coloring methods to other contents, such as human, animals, etc. (based on the conclusion of enabling high-definition repair in sd-webui).

The generated material's styling and colors have been greatly improved. For example, it can capture the rationality of combining the keyword smile with the treasure chest icon, which would have been much inferior in the SD1.5 version.

The above results were generated in the comfyui environment. The shortcomings include inaccurate fine details in styling and dark-colored outputs. Running the model on sd-webui with high-definition repair enabled would yield better results.

Please use the base_xl official original model in conjunction with the lora weights controlled between 0.6-0.8. Trigger words to use: Trilightlabs style, followed by icon attributes such as: gem, stone, drink, ring, shoes, flower, treasure chest, gift box, money, food, fruit, gloves, sword

This set of icons covers almost all types of game icon categories needed in daily use. You can try various attribute keywords. The style of this lora model leans towards the realistic Japanese and Korean style of game graphics.