Trigger: natasha romanoff
Extra prompts: party outfit, bodysuit (her Avengers uniform), messy hair
This model was trained to be used exclusively with Illustrious Personal Merge (v3.0): https://tensor.art/models/793160372556268412/Illustrious-XL-personal-merge-V3.0
It may work on different models (such as NTR Mix), but most likely won't, especially not with realistic Illustrious finetunes.
Since this LoRA was trained on top of an anime model it has certain caveats:
- Generic quality improving negative prompts will reduce likeness and realism to a certain extent. I suggest leaving the negative prompt empty unless you want to remove something from your gens.
If you are having trouble with bad anatomy or composition you could try reducing lora strength a bit and if nothing else works try adding this as negative prompt: lowres,worst quality,low quality,cropped,jpeg artifacts,monochrome, greyscale, anime, 3d
Much more than that and likeness will probably be severily impacted. Please let me know if you find a negative prompt that improves overall quality without sacrificing realism or likeness.
- Some (but not all) generic quality improving prompts may also reduce likeness. I've found that the following will improve gen quality without impacting the lora: (masterpiece,best quality,newest,absurdres,highres)
- Adding (photorealism) as positive can help improve overall quality.
I suggest Euler A/Karras, CFG 5, 35 steps, but feel free to experiment and come up with better settings.
Strength between 0.85 and 1. I think 0.9 is the sweet spot. You can also use Lora Block Weight on lower steps (1 up to step 5) without much negative impact if the gens are deviating too much from the main model.
Should use ADetailer for improved likeness
ATTENTION: This was trained as a DoRA, if your local install (A111/Forge/Comfy/etc) is a few months old it might not be compatible. In this case just update your install and the model should load properly.