Negative used in training: 'this was used as negative when training the lora'

//--//

All training images begin with the phrase 'this image is a digital illustration by esuthio'

//---//

For Image captioning. I suggest using this Gemini 2.5 custom caption prompt (found in TA LoRa trainer) : begin the descripton with the text 'this image is a digital illustration by esuthio' . Then itemize the things you see in this image. Avoid using bold font. Avoid using newline

//--//

LoRa training include

• Deific robots with skull motifs by Esuthio (main theme)

• Tsutomo Nigei Blame aesthetics

• Norihiro Yagi Claymore faces

• Brian Sum robot concept art

• LanaRPG pixel sprite characters

• Text training in every frame of short words

125 images x 8 repeats x 40 epochs = 40,000 training steps

At rank 32 , alpha 32 , learning rate 0.0003 using 'cosine with repeats'

Training data https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/esuthio_training_data.zip