Stage aware method erases concepts in text to 3D models
STAGE: Subspace-Targeted Affine Generative Erasure for Text-to-3D Models
Computer Vision and Pattern Recognition
Summary
Sometimes when making 3D images from text, creators want to remove certain shapes or textures without messing up everything else in the model. The authors show that changes to shape and object details need to happen in an earlier stage of making the 3D model, while changes to appearance or materials happen later. They introduce a new way called STAGE that carefully edits just the right parts at the right times without retraining the model. STAGE works better than previous methods in forgetting unwanted parts while keeping everything else intact.
What this means in practice
- •For 3d content creators: Remove specific unwanted shapes or textures from generated 3D models while preserving other aspects effectively.
- •For virtual environment developers: Edit 3D model components in text-to-3D pipelines to customize scene elements without full regeneration.
Authors
Karol Dziekan, Przemysław Spurek, Dawid Malarz
Abstract
Concept erasure suppresses a target concept while preserving behavior on unrelated inputs. Existing closed-form methods were designed for 2D image diffusion and assume a single generative pathway, so one edit must cover geometry and texture at once. Native 3D generators, which synthesize structured 3D representations directly rather than by lifting 2D samples, violate this assumption. We show that shape and object concepts must be erased in the structural stage of the pipeline and material concepts in the appearance stage. We therefore formulate erasure in native text-to-3D as a stage-aware editing problem and introduce STAGE, a training-free, closed-form framework. STAGE confines each edit to the low-dimensional subspace spanned by the differences between erase and anchor embeddings, and relaxes the norm-preserving (orthogonal) constraint of prior editors into a least-squares affine correction that maps target activations onto safe anchors subject to a penalty on the displacement of retained prompts. The correction applies to the structural stage, the appearance stage, or both. We find that the stage an edit must reach is determined by concept type. On TRELLIS, the standard open native 3D generator, across 15 shape, material, and object concepts, STAGE reaches 66.7 on a composite score that balances forgetting the target concept against preserving everything else, aggregating CLIP-based semantic and physical metrics, versus 53.2 for the strongest adapted baseline. Code: https://github.com/gmum/STAGE/ Project Page https://gmum.github.io/STAGE/