jewel-model-overfit

A full fine-tune of the FLUX.2 [klein] base 4B transformer for image-to-image jewellery retouching. It was deliberately overfit on the 12-pair jewel-overfit-dataset for 5000 steps (final loss ≈ 0.095, lr 1e-5).

Only the transformer is included here. Load the VAE, text encoder and scheduler from the base model.

• Prompt: studio-retouch the jewellery
• Weights: transformer/diffusion_pytorch_model.safetensors (bf16, ~7.7 GB)

import torch
from diffusers import Flux2Transformer2DModel

transformer = Flux2Transformer2DModel.from_pretrained(
    "itsmohit/jewel-model-overfit", subfolder="transformer", torch_dtype=torch.bfloat16
)

License: follows the FLUX.2 [klein] base 4B license.