Gate3DGen: A Pre-processing Model that Selectively Purifies Real-World Images for 3D Generation

Project Page · Code

Gate3DGen prepares real-world image inputs for existing 3D generators. It routes inputs classified as clean through a KEEP branch and sends the remaining inputs to a PURIFY branch. The downstream 3D generator stays frozen.

  • Degradation-aware router: predicts KEEP/PURIFY and provides features to condition purification (finetuned/gateb).
  • 3D-oriented one-step purifier: uses the router's features to purify degraded inputs while preserving object identity, contours, and textural cues (finetuned/restoration).

Checkpoints

This release contains FP16 replacement tensors (about 16.3 GB) and their configurations and manifests. It requires reconstruction against FireRed-Image-Edit-1.1, revision 3bc3f2a12722fd9883eb6357500de191d56baaf5. Stored tensors replace the corresponding base tensors; unchanged base tensors are cast to FP16.

These files are not the assembled full-checkpoint layout described in the code repository's download guide. The FP16-converted weights differ in precision from the original training checkpoints, and this release does not include a new image-quality evaluation. Upstream model terms apply.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for hllee0615/gate3dgen

Finetuned
(7)
this model