Modality Forcing: Unified Spatial Generation via Per-Modality Noise Control
Modality Forcing enables scalable joint image-depth generation by assigning independent noise schedules to separate modalities within a single DiT, allowing training on sparse real-world depth data without dense annotations. The approach scales predictably with model size and data—larger models produce more accurate depth predictions—and achieves 57% relative improvement in AbsRel over existing joint generative models while matching specialized monocular depth estimators. This demonstrates that image generation pre-training is a viable scalable objective for spatial perception tasks.