Efficient inference · First-author research / 2026
Retire3D
Selective computation for 3D generation
The question
In coarse-to-fine 3D generation, can high-resolution computation be allocated more selectively while retaining the context needed by the model? Retire3D investigates this question without retraining the underlying generator.
My contribution
I led a training-free approach that selectively allocates high-resolution computation while preserving the full latent representation and attention context. The work evaluates Direct3D-S2 and TRELLIS.2.
Evaluation
The submitted manuscript reports a 1.528× end-to-end speedup over native TRELLIS.2 when Retire3D is combined with temporal caching. This number describes that combined configuration and baseline.
Current status
Retire3D is a first-author manuscript submitted to ICLR 2027.
Want to talk about this work? Get in touch ↗