physical violence

weapon violence

wide hips

revealing clothes

thick thighs

downblouse

convenient censoring

huge breasts

pg-13

corpses

suggestive

oral invitation

pg13

sexy

sexual situations

male nudity

disturbing

male swimwear or underwear

female swimwear or underwear

partial nudity

undressed

female nudity

breasts out

exposed female nipple

breast out

lingerie

male underwear

hair over breasts

female swimwear

gigantic breasts

no panties

graphic violence or gore

covered nipples

huge butt

strapless leotard

sitting on face

emaciated bodies

one breast out

nsfw

female underwear

nude

graphic male nudity

adult toys

illustrated explicit nudity

nudity

graphic female nudity

hentai

futanari

porn

sexual intent

genitals

peeing

vore

oral

sexual activity

anal

blowjob

dildo riding

incest

hanging

hate symbols

nazi party

white supremacy

diapers

scat

self injury

hate speech

urine

extremist

child on child

latex clothing

swimwear

bukkake

fellatio

cumshot

implied fellatio

eat_cum

cumdrip

cum in pussy

cum on face

after fellatio

cum on hair

cum on body

cum on tongue

cum on hands

cum in mouth

triple fellatio

autofellatio

fucked silly

cum on pussy

LoRa, which I trained on my own computer, is still working on how to adjust parameters and improve the data set.The current version is 0.1, waiting to be fully perfected.

Trained with a combination of English tags and natural Chinese language generated by an LLM, for a total of 32 epochs and 10624 steps. The flow matching loss decreased from 0.37 to 0.341..., but the ideal range should be 0.22~0.23, with a very good value being 0.1~0.12.Based on my training experience with qwen-image and flux, which I have not uploaded to Civitai as they are company assets, the issue is with the prompt structure.Currently, I need to find a good prompt structure to solve the training problems, but this is difficult. Both pure natural language and pure tags have inherent flaws, and they should be integrated into a better, standardized structure. The quality of the Chinese prompt captions is poor, which is why the convergence effect is not very good...From my past experience, finding a good prompt and structure can speed up the convergence rate by several times while maintaining overall quality, but this exploration takes a very long time.Regarding some overfitting issues, I suggest lowering the weight to 0.8~0.9 or even lower for better results. I believe that the current base model's fine-tuning is still hampered by a poor prompt structure, which leads to insufficient training. Of course, the VAE is also a problem to some extent.If there were enough funds, we would need to fine-tune millions of images with their corresponding prompts. This would require about eight to sixteen epochs to correct, and the engineering effort would be massive, requiring strong personnel coordination, which is very difficult. The learning rate could be appropriately lowered for directional guidance. (1e-4 with large batches -&gt; 5e-5 or lower with large batches).

NetaYumev2plus_lumina_0.1

Retrain based on IL 1.0, and use more consistent tags to improve the training effect.Reclean and create synthetic data to improve quality and better preserve details.It is recommended to use WAI-NSFW-illustrious-SDXL v14.0 output to alleviate some overfitting problems.Since the details of various combinations are not compared using magnified details, the headdress and bow on the sansochan wedding dress may not be effectively eliminated on the back and are easy to generate.It supports some NSFW but not many, and the total number of training samples including original appearance is about 150.When closer to the face or upper body, eye detail is better.

v0.3

Some dataset improvements were made and lower learning rates and larger amounts of synthetic data were used, but the results were not good enough to accurately match all the characteristics of the characters.Fine-tuning models using other illustrious models do not adapt well.I am now considering training other comic characters and various elements to improve my related experimental data and cognition.

v0.2

v0.1

Sansochan

擷取.JPG