Modality Forcing for Scalable Spatial Generation: Joint Image-Depth Generation with a Single DiT

Forum topic · 小凯 · 2026-06-13

Summary

This forum post summarizes the arXiv paper "Modality Forcing for Scalable Spatial Generation" (arXiv:2506.10667) by Bardienus Pieter Duisterhof, Deva Ramanan, and Jeffrey Ichnowski. The paper proposes Modality Forcing, a simple and scalable post-training method for joint image-depth generation using a single Diffusion Transformer (DiT) trained on sparse depth data. By assigning separate noise levels per modality, the method enables conditional and joint generation of images and depth in any permutation. Per-modality decoders allow training on sparse real-world depth and yield strong, generalizable depth prediction. The authors show the approach inherits the scalability of text-to-image pretraining: training T2I models from 370M to 3.3B parameters, larger models trained on more image data produce more accurate depth. Their strongest model competes with state-of-the-art monocular depth estimators and reduces AbsRel by 57% versus existing joint image-depth generation models, suggesting image generation is a scalable pretraining objective for spatial awareness.

Paper Overview

Field: Computer Vision Authors: Bardienus Pieter Duisterhof, Deva Ramanan, Jeffrey Ichnowski Published: 2025-06-13 arXiv: 2506.10667

Abstract (Translated)

Text-to-image (T2I) models contain rich spatial priors. Synthesizing photorealistic, cluttered scenes requires an understanding of geometry, including perspective and relative scale. Prior works adapt T2I models to leverage this prior for depth prediction, but they require dense depth data and involve complex recipes. We propose Modality Forcing, a simple, scalable post-training recipe for joint image-depth generation using a single DiT trained on sparse depth data. Modality Forcing enables conditional and joint generation of image and depth in any permutation by assigning separate noise levels per modality. Per-modality decoders let us train on sparse, real-world depth and achieve strong, generalizable depth prediction.

We further show that Modality Forcing inherits the scalability of T2I pretraining: by training a suite of T2I models from scratch (370M to 3.3B parameters), the authors find that larger models trained on more image data yield more accurate depth. Their strongest model is competitive with state-of-the-art monocular depth estimators, reducing AbsRel by 57% relative to existing joint image-depth generation models. These results provide strong evidence that image generation is a scalable pretraining objective for spatial awareness.

Key Takeaways

Modality Forcing: a simple post-training recipe enabling joint and conditional image-depth generation in any permutation via per-modality noise levels.
Sparse-data friendly: per-modality decoders allow training on sparse, real-world depth annotations.
Scalability: experiments across 370M–3.3B parameter models show depth accuracy improves with model size and data, inheriting T2I scaling laws.
Results: competitive with state-of-the-art monocular depth estimators; 57% AbsRel reduction versus prior joint image-depth generative models.

---

*Auto-collected on 2026-06-13*