OVB FLUX.2-dev checkpoint 500

This repository contains the transformer weights from optimizer step 500 of the OVB FLUX.2-dev fine-tuning run flux2_dev_v2_no_masks_32node_global512_6250_s42.

This is an early intermediate research checkpoint, not the final trained model. The run used 32 H200 nodes, effective global batch 512, seed 42, and disabled training validity masks. Only the fine-tuned transformer is included; the VAE, text encoders, and other base-pipeline components are not duplicated.

Loading

Access to the gated black-forest-labs/FLUX.2-dev base model is required.

import torch
from diffusers import Flux2Pipeline, Flux2Transformer2DModel

transformer = Flux2Transformer2DModel.from_pretrained(
    "valentinospariza/ovb",
    torch_dtype=torch.bfloat16,
)
pipe = Flux2Pipeline.from_pretrained(
    "black-forest-labs/FLUX.2-dev",
    transformer=transformer,
    torch_dtype=torch.bfloat16,
)

License

These weights are a derivative of black-forest-labs/FLUX.2-dev and are provided under the FLUX Non-Commercial License. Users are responsible for complying with that license and the base model's acceptable-use requirements.