Modelpublic

uviol/scat-flux-2-klein-9b-v1

sourceHugging Faceapache-2.0updated 7mo agoView on Hugging Face

3likes333downloads

Model Card

Scat LoRA — FLUX.2-klein-9B

Model description

Scat LoRA. Use natural language, "shitting" was the preferred verb. Only trained on photos so far, style LoRAs are probably necessary for non-realism.

Training info

v1

Dataset

Similar dataset to Qwen-Image-2512 version but cut down slightly (102 images) and recaptioned with Qwen3-VL-235B-A22B-Instruct-heretic. Caption quality doesn't seem significantly better than Qwen3-VL-32B-Thinking-heretic. Will try Qwen3.5 next time.

Hyperparameters

Parameter	Value
Base model	FLUX.2-klein-base-9B
Trainer	musubi-tuner
Optimizer	AdamW8Bit (defaults)
LR Schedule	cosinewithmin_lr
LR	1.5e-4
Min LR ratio	0.1
Lora+ LR ratio	4
Warmup steps	300
Epochs	200 (20400 steps)
Timestep Sampling	flux2_shift

Evaluation

Fidelity is higher than Qwen-Image-2512 version but Klein's understanding is weaker and can't produce as complex poses and scenes. Frequently chooses the wrong hole. Frequent anatomy errors, but no worse than the base model. Genital quality is not great without a supplementary LoRA like SNOFS.

Download model

Download