lainflux-klein

NOEMA · part of the klein family — character & style LoRAs on FLUX.2 klein (4B).
Run it privately at noema.art — no email, pay anonymously.

FLUX.2 [klein] 4B conversion of ms2stationthis/lainflux (originally FLUX.1-dev). Trigger: "lain".

Trigger word: lain Retrained onto FLUX.2 [klein] 4B from ms2stationthis/lainflux (FLUX.1-dev).

Sample Outputs

lain, This is a digitally rendered image featuring a small, anime-style girl, referred to…	lain, This is a digital artwork depicting a young lain with a pale skin tone and short, d…
:---:	:---:
lain, This is a digital anime-style illustration depicting a young girl, lain, with a fai…	lain, This image is a digital drawing in a stylized anime art style. It features the youn…

Usage (Diffusers)

import torch
from diffusers import Flux2Pipeline

pipe = Flux2Pipeline.from_pretrained(
    "black-forest-labs/FLUX.2-klein-4B", torch_dtype=torch.bfloat16
).to("cuda")
pipe.load_lora_weights("ms2stationthis/lainflux-klein")
image = pipe("lain, a character portrait", guidance_scale=4.0, num_inference_steps=25).images[0]
image.save("out.png")

Recommended Settings

LoRA strength	Guidance	Steps	Resolution
0.8–1.0	4.0	25	1024×1024

Training Details

• Base: black-forest-labs/FLUX.2-klein-base-4B
• Steps: 4000 · Network: LoRA rank 32 / alpha 32
• Optimizer: adamw8bit, lr 1e-4 · Scheduler: flowmatch
• Resolution: 512, 768, 1024 (multi-res bucketed) · Precision: bf16 train / fp16 save

Reproduction

This repo includes the full dataset/ (20 image-caption pairs) and the exact config.yaml so the LoRA can be retrained as-is.

About

NOEMA is a privacy-by-construction generative studio. Run this model — and 80+ others — privately at noema.art: no email to start, pay anonymously, go fully private anytime.

_{NOEMA · a complete studio, completely private · noema.art}