ShapeUp

Weights for ShapeUp, image-conditioned 3D shape editing built on Step1X-3D.

Given a source shape and an edit image, ShapeUp generates the edited mesh with a rectified-flow DiT conditioned on (a) the DINOv2+CLIP embedding of the edit image and (b) the VAE latents of the source shape.

Repository layout

geometry/
└── adapter-1024-550k/          # geometry adapter (this release)
    ├── adapter_config.json
    └── adapter_model.safetensors
texture/                        # texture weights (coming soon)

geometry/adapter-1024-550k

Only the trained adapter is stored here — everything else loads from stepfun-ai/Step1X-3D. The adapter is the complete set of trainable parameters:

LoRA on the DiT (rank 128, alpha 128)	288 tensors
proj_shape_condtion (source-shape projection)	3 tensors
total	291 tensors / 75.6M params

Trained with n_part_latents=1024, 550k steps.

adapter_config.json carries a manifest (num_tensors, num_elements) so the loader can confirm it received the whole adapter.

Usage

from step1x3d_geometry.models.pipelines.pipeline import ShapeUpGeometryPipeline

pipe = ShapeUpGeometryPipeline.from_config(
    "configs/train-geometry-diffusion/"
    "step1x-3d-geometry-shapeup-lr1e-parts-reconstruction-train-new-data-multistep-inference.yaml",
    adapter_path="Inbar2344/ShapeUP",              # or a local directory
    adapter_subfolder="geometry/adapter-1024-550k",  # this is the default
    device="cuda",
)

out = pipe(
    {"surface": surface, "sharp_surface": sharp_surface, "image": [edit_image]},
    seed=0,
    num_inference_steps=30,
    octree_resolution=256,
    cfg_type="separate_visual",
    guidance_scale_visual=2.5,
    guidance_scale_shape=3.5,
    output_type="trimesh",
)
out.mesh[0].export("edited.glb")

The base Step1X-3D weights (VAE, DiT, DINOv2+CLIP encoder) are downloaded automatically from the Hub on first use.