(LECO) Erase speech bubble and sound effects for noobaiXLNAIXL_epsilonPred10

The model noobaiXLNAIXL_epsilonPred10Version.safetensors is an excellent one, but when generating NSFW content, speech bubbles and sound effects may sometimes appear.

To address this, I used LECO to remove them.

Of course, these elements can be avoided by including them in the negative prompt during generation. However, I created this adjustment for those who might want to merge models as well.

Although I haven’t tested it, I believe it should also work with derivative models.

Generally, setting the value to around 4.0 should almost entirely remove these elements, but please fine-tune it as necessary.

The training time is approximately 4 hours. Feel free to use this if you want to avoid reinventing the wheel.

settings

name_or_path: "noobaiXLNAIXL_epsilonPred10Version.safetensors"

rank: 64

alpha: 0

lr: 1e-4

optimizer: "AdamW"

lr_scheduler: "cosine"

per_steps: 1000

target: ""

positive: "speech bubble, sound effects"

unconditional: ""

neutral: ""

action: "erase"

guidance_scale: 1.0

resolution: 1024

dynamic_resolution: false

batch_size: 2