Lora with Jeong Somin for Flux. FLUX.1-dev base, 4e-4 learning rate, bf16 data type. 100 epochs, 2000 learning steps (trained up to 4000, but I think it's too much). Trained in Colab. Dataset: 20 photos, 768x768 resolution, descriptions like "Somin, woman, 1girl, long hair, breasts, looking at viewer, brown hair, dress, cleavage, brown eyes, jewelry, upper body, earrings, necklace, off shoulder, black dress, lips, makeup, realistic, off-shoulder dress" were added to each photo. Of the jambs, "Somin", "woman", "1girl" can be used as trigger words, and the inscription "Merci" appears in the background. This is my first lore, so yes, such shortcomings.

I use Swarm UI. Steps 20, CFG Scale 1, Sampler Euler, Scheduler Beta, Flux Guadance Scale 3.5. Model flux_dev_fp8_scaled_diffusion_model.