DrivingGFM v1.0

Generalist Foundation Model for Unified Driving Visual Perception — DrivingGFM Team, Tongji University

Quick start

pip install diffusers transformers accelerate torch

Recommended entry point, using the DrivinggfmPipeline wrapper shipped within this repository. trust_remote_code=True is required to load the custom pipeline code from the repository.

import torch
from PIL import Image
from diffusers import DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained(
    "ZehongZhu/drivinggfm",
    torch_dtype=torch.float16,
    trust_remote_code=True
)
pipe.to("cuda")

out = pipe.drive(
    Image.open("scene.png"),
    "make it a rainy night road scene",
    num_inference_steps=20,
    image_guidance_scale=1.5,
    guidance_scale=7.5,
    seed=2026,
)
out.save("result.png")

License

CreativeML Open RAIL‑M, full text in LICENSE. Commercial use and redistribution are permitted, provided the full license text and attribution are retained, downstream users are bound by the Attachment A use restrictions, and modified files are clearly marked as modified.