WILD AND WET

作者：Franklin_Sports

This is a StyleLoRA model trained on sports and fashion magazines and advertisements published in the United States from 1989 to 1994. I trained it using a dataset of 500 images I created by cutting up and scanning magazines I actually had at home. Much of the training was based on swimsuit pin-ups. Those wild, humid beaches of that era...

The main trigger word is “WAW90.” It broadly reflects the swimwear and appearance of the people in the training data. However, it cannot reproduce specific individuals. You can recreate a 1990s-style texture and atmosphere by using prompts that directly specify the era and image quality, such as “A vintage 1990s pinup photo.” Additionally, increasing the LoRA strength enhances the image quality and texture typical of 1990s magazines. It appears to function with a strength of up to 2.0. Since the images were scanned from magazines, most of the training data contained captions or text in locations that are difficult to separate. If the prompt lacks sufficient details about the background, or if the LoRA strength is high, captions not included in the prompt may appear in the generated image. To avoid this, you can either write a detailed prompt describing the background, lower the LoRA strength, or remove them later using the edit tool. Alternatively, you can come up with your own captions and include them in the prompt. If you’re using “raw” instead of Krea2Turbo, using negative keywords to exclude captions is another option. I’ve posted numerous sample images and prompts, so please feel free to use them as a reference. I hope I’ve managed to capture the atmosphere of early ’90s magazine spreads.

The sample images posted here are primarily generated using ComfyUI. The checkpoint and sampler settings are as follows: Checkpoint: krea2Turbo or Krea2 turbo NSFW AIO Guidance scale: 1 Steps: 8 Sampler: Euler In some cases, images generated on Civitai or those using other checkpoints may be included. Note that negative prompts do not work with krea2Turbo; while some image metadata may contain negative prompts, these were simply copied and pasted from my workflow presets and are not actually functioning.

Ver 1.1： I have rebuilt the dataset and retrained the model. The initial version of the dataset included magazine advertisements. While these were effective for capturing the fashion and atmosphere of the late 1980s and early 1990s, they also contained a significant amount of text. Since Krea2 is a highly intelligent model, it typically avoids incorporating such text into generated images; however, when using prompts like "A vintage 1990s magazine pinup photo"—intended to recreate the style of 1990s-era pin-ups—the text would sometimes reappear. While such issues can usually be addressed via negative prompts, Krea2 Turbo does not support them. Although one could use custom nodes or the "raw" setting, I wanted to improve the model's overall usability, so I refined the dataset to include only images free of text. I also incorporated new source material, resulting in a final dataset of 500 images used for training.