Instructions to use fassabilf/sd15-ift-us-gender with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusers
How to use fassabilf/sd15-ift-us-gender with Diffusers:
pip install -U diffusers transformers accelerate
import torch from diffusers import DiffusionPipeline # switch to "mps" for apple devices pipe = DiffusionPipeline.from_pretrained("fassabilf/sd15-ift-us-gender", dtype=torch.bfloat16, device_map="cuda") prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k" image = pipe(prompt).images[0] - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- Draw Things
- DiffusionBee
import torch
from diffusers import DiffusionPipeline
# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("fassabilf/sd15-ift-us-gender", dtype=torch.bfloat16, device_map="cuda")
prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
image = pipe(prompt).images[0]SD 1.5 β IFT gender US (checkpoint epoch 5 / 10 / 15)
Supervised fine-tuning Stable Diffusion 1.5 di atas dataset IFT
fassabilf/ift-train-us-sd15 (1400 train + 308 val, 14 okupasi US, sumbu gender).
Caption yang dilatih netral (tanpa kata atribut); distribusi gender-nya yang
dikoreksi lewat komposisi gambar per okupasi.
Tiap direktori pipe_epN/ adalah pipeline diffusers utuh dan bisa langsung dipakai:
from diffusers import DiffusionPipeline
pipe = DiffusionPipeline.from_pretrained("fassabilf/sd15-ift-us-gender",
subfolder="pipe_ep5")
safety_checker sengaja tidak disertakan: ia menghitamkan sebagian keluaran dan
itu menggerus distribusi gender yang justru sedang diukur.
Setup training
| Kode | train.py --sft dari Diffusion-DPO (Wallace dkk.) |
| Yang dilatih | UNet penuh (860M); VAE + text encoder beku |
| Presisi | bf16, gradient checkpointing |
| Resolusi | 512, tanpa horizontal flip |
| Batch | 8 Γ grad-accum 4 = efektif 32 |
| Optimizer | AdamW, lr 1e-5, constant_with_warmup, warmup 50 |
| Langkah | 660 = 15 epoch (44 optimizer-step/epoch) |
| Checkpoint | tiap 220 step = tiap 5 epoch |
| Hardware | 1Γ RTX 5090 32 GB, VRAM puncak ~20,3 GB |
| Waktu | 18 menit 56 detik |
Peringatan
Angka evaluasinya belum final saat kartu ini ditulis. Sinyal awal menunjukkan fine-tuning ini mempolarisasi, bukan mengkalibrasi: okupasi bertarget tinggi terdorong melewati targetnya, okupasi bertarget rendah terdorong makin ke laki-laki. Jangan pakai checkpoint ini sebagai model debias sebelum tabel MAE final dibaca.
- Downloads last month
- -
Model tree for fassabilf/sd15-ift-us-gender
Base model
stable-diffusion-v1-5/stable-diffusion-v1-5