fassabilf/sd15-ift-uk-real

SD 1.5 IFT — foto asli, UK, gender

UNet hasil train.py --sft di 400 foto asli Pexels (2000 baris train / 600 val, dataset fassabilf/ift-train-uk-real), 20 epoch, --seed 0.

Arm UK: 20 okupasi terlatih (mechanic, civil_engineer, engineer, engineering_technician, it_support, caretaker, it_operations, sales_executive, lab_technician, finance_analyst, sports_coach, higher_ed_teacher, director, accountant, further_ed_teacher, secondary_teacher, office_manager, nurse, teaching_assistant, personal_assistant; sufiks _uk), ground truth ONS, prompt "in the United Kingdom". 63 step/epoch (US: 44) — jangan menyamakan epN UK dengan epN US. Lima okupasi held-out (tidak pernah dilihat waktu training) dievaluasi di ep5/ep20: metal_fitter, senior_care_worker, care_worker, crossing_patrol, education_support. Hasil paired ΔMAE_real vs base SD 1.5 (20 okupasi, cluster bootstrap 95 %): ep5 −0,044 [−0,090, −0,003], ep15 −0,055 [−0,114, −0,002], ep20 −0,057 [−0,099, −0,018]; ep10 memuat 0. Detail di analysis/REPORT.md.

VAE, text encoder, dan tokenizer identik dengan SD 1.5 base — tidak ikut diunggah. Rakit pipeline-nya dengan memasang UNet di sini ke stable-diffusion-v1-5/stable-diffusion-v1-5 (lihat scripts/ift/ckpt_to_pipeline.py di repo kode).

Checkpoint yang diunggah: ep5, ep10, ep15, ep20. Setelan lengkap di hparams.yml; angka evaluasi, error bar, dan lantai MAE di analysis/.

Model ini artefak penelitian tentang bias distribusi gender, bukan model produksi.