Slim Studio | MOSS-Image LoRA

Nothing in common with Influencergram. Influencergram was build using way more diverse dataset with different body types. This one knows only one and knows it very well.
Both are using completely different datasets.

What's different

Uncensored captions. Every training image was captioned with an uncensored captioner, so nudity, poses and anatomy are named plainly in the data. A lot of InfluencerGram's captions were "safe" and something could fail understanding some NSFW prompts. This is the fix. Or not i really don't know - could be better, could be worst at some that Influencergram's good at.

One body type. Slim, skinny, and petite. Every time. No mix of builds pulling it in different directions.

Small and hand-picked. 270 images kept out of 700+ used for Influencergram, only 2 to 5 per person, so it learns the look, not faces. Yet it'll still drift towards more asian look at higher strengths.

How I use it

MOSS-Image SFT + 8-step DMD2 LoRA @ 1.0

er_sde / simple, 8 steps, CFG 1

832x1248

Strength 1.0 to start. Tested from 0.7 to 1.4, the strength sheets in the gallery show the range.

Prompting

Describe her like a photo caption: body, outfit (or none), pose, where she's looking, the camera (selfie, mirror selfie, someone else shooting), the room and the light. The showcase images have their full prompts.

Training

Base: MOSS-Image Base int8 (use it on SFT)

270 images, 4000 steps (about 15 epochs), batch 1, 1024 buckets

Rank 32 / alpha 32, AdamW, lr 1.7e-4, constant with warmup

Trained and captioned locally on an AMD RX 9060 XT

Also by me: InfluencerGram, for the influencer feed look.