She-Hulk Image Model with LoRA
Description:
This LoRA model is capable of generating both safe-for-work (SFW) and not-safe-for-work (NSFW) versions of the Marvel character, She-Hulk, from the MCU TV Series.
Remember to also add "green skin" along side the trigger words to the prompt for better results, as sometimes it might generate good looking She-Hulk, just in normal Non-Green Color Skin.
Important Instructions:
Before utilizing this model, please take a moment to read these essential guidelines:
Training Process: Initially, attempts were made with a lower training step count, but due to the intricacies of She-Hulk's character design, satisfactory results were elusive. Consequently, the model was trained on a more extensive dataset, comprising precisely 337 images. However, this extensive training has led to some degree of overfitting.
Adjusting Output: To optimize the model's performance, adjust the value following the ":" symbol in <lora:SHEHULK_SDXL-000005:1>. A recommended setting is 0.75, but you have the flexibility to fine-tune this value based on your preferences and desired results. For example, if you wish to alter the style to resemble anime using other style LoRAs or Checkpoints, decrease accuracy and increase flexibility by using a value of 0.65. Values between 0.65 and 0.75 are generally ideal, with occasional usage of 0.6 or 0.8 for specific outcomes. The provided example images demonstrate the model's precision and adaptability.
Adjusting Prompt: As the model is a bit overfitted, you might need to emphasis on some words.
For example, you can see 1st and 2nd image, they have every parameter the same except for her hair. But I needed to put a bit more emphasis on "tied back hair" as compared to "long hair". You can see the examples if you do not know how to increase the emphasis of the word in the prompt. Also same can be done on words in Negative Prompt.
Resolution: While the model was extensively tested setting at a resolution of 1000x1200 (so also the recommended), it is adaptable to various resolutions, consistently delivering impressive results.
Optional Tip: For further refinement, consider using a lower resolution and apply Hi-Res Fix at 1.5x or 2x with denoising set between 0.3 to 0.4 and 20-30 steps.
Model Size: The model's larger size is intentional. While it can be downsized using Kohya_ss, increasing its size is not feasible. Downsizing may impact image quality, and as of now, there's no workaround for this limitation.
For those unfamiliar, reducing the model's size may compromise the output quality. Therefore, it is advisable to work with the provided larger model for the best results.
If you have read till here, I would like to point out this is my first model training, so if it works as you want it and the results are good, kindly rate the model.
Thank You !!!
