Lora keyword is the negative prompt: 'this was used as negative when training the lora'

//---//

E39 has better stability than E41

Huggingface repo (my own) : https://huggingface.co/codeShare/Drills3D_Chroma_LoRa

//---//

Make sure to use 't5xxl_fp16.safetensors' for long prompts

//---//

This LoRa is aimed at creating female characters in indoor spaces. With focus on:

• Candid movement (face close ups , arm / leg movement , using characters from KawaiiDetectiveEnthusiast and Drills3D)

• Interior design (furniture / blur / depth)

• Skin tone (3DCG - suitable for posting on regular SFW Tensor as well as Tensorhub)

• Liminal spaces (Pinterest photos)

• Manga frames (Claymore and Tsutomo Nihei Blame! mixed)

//--//

Training data are still frames from 3DCG animation into 4x1 sliced 1024x1024 composites.

//--//

250 image composites x 4 repeats x 40 epochs

Rank 32 Alpha 32 at LR= 0.00035 using cosine with repeats

No bucketing , using Kohya trainer

//---//

Suggested usage: Do Gemini 2.5 caption of 3DCG or photo , including some danbooru tags into custom prompt.

Use t5xxl_fp16.safetensors text encoder or similiar (needs to be manually selected upon remixing post due to site bug)

//--//

I remix SFW lewd Chroma prompts from time offsite to time and decided to make this LoRa to make the remixes more presentable.

//---//