Sign In

Lifesim Bachelorette Portrait Style

53
302
221
13
Verified:
SafeTensor
Type
LoRA
Stats
302
221
Reviews
Published
Jan 18, 2024
Base Model
SD 1.5
Training
Steps: 3
Epochs: 10
Usage Tips
Clip Skip: 1
Strength: 1
Trigger Words
lifesim_portrait_style
<lora:lifesim_portrait_style-10:1>
solo
1girl
looking at viewer
portrait
Hash
AutoV2
3B79A0BADC

I'm excited to share my first LORA creation, inspired by this guide from @holostrawberry. This model generates pixel art female portraits of NPCs in games like Stardew Valley. Here's a breakdown of recommendations for optimal use:

  • Dataset Details: The LORA is trained on approximately ~400 female portraits, originally in 128x128 pixel format. These images are sourced from a handful of different life and farming sim games, not just SDV, to give it some flexibility.

  • Upscaling Method: To maintain the original pixel art style, each image was upscaled to 512x512 using nearest neighbor sampling.

  • Tagging: I used WD1.4 for tagging. While I've made an effort to refine and prune tags for accuracy, the curation process was brief. I encourage users to explore and adjust tags as needed.

  • Optimal Weight Settings: For general use, set the weights between 0.7 to 1. Lower weights tend to yield better results for NSFW content.

  • Additional Tools for Enhanced Pixelization: For authentic pixel art results, pair this LORA with any pixelization extension. Downsampling to a 4x scale is recommended to achieve crisp, clear pixel art outputs.

  • Activation Tag: Utilize the 'lifesim_portrait_style' activation tag alongside the LORA weight. This combination is crucial for achieving the intended style. Other recommended, but optional, tags are '1girl, black background, portrait, upper body, looking at viewer, pixel art'

  • Prompt Guidance: Simplicity is key. Avoid overcomplicating your prompts. Refer to the provided examples for guidance on crafting effective prompts.

  • Training and Compatibility: The model was trained on the base SD 1.5. For optimal results, use it with anime-based checkpoints, which complement the style and enhance the output quality.

Regarding the XY plot at the end:

  • X-axis: Epochs 1 to 10 (left to right)

  • Y-axis: Weights 0.7 to 1.2 (top to bottom)

I don't have any plans to update this model in the future, it was a one time experiment, but feel free to do whatever you want with it and remember to support the creators and artists of such games.