README.md

metadata

title: Modality Forcing
emoji: 🏢
colorFrom: blue
colorTo: gray
sdk: gradio
sdk_version: 6.16.0
python_version: '3.12'
app_file: app.py
hardware: zero-h200
pinned: false
license: apache-2.0
short_description: Text → RGB + depth + 3D point cloud.

Modality Forcing for Scalable
Spatial Generation

Bardienus Pieter Duisterhof^1,2 · Deva Ramanan¹ · Jeffrey Ichnowski¹ · Justin Johnson² · Keunhong Park²

¹ Carnegie Mellon University    ² World Labs

Preprint, 2026

Modality Forcing turns a pretrained text-to-image model into a joint image–depth generator with a simple post-training recipe.

Overview

This Space hosts the interactive demo of Modality Forcing: joint text → RGB + depth diffusion built on FLUX.2. A single DiT supports every permutation of conditional and joint generation by assigning a separate noise level to each modality:

Mode	Input	Output
Joint	text prompt	RGB + depth + 3D point cloud
Image → depth	text + image	depth + 3D point cloud (any aspect ratio, letterbox resize)

Abstract

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 pre-training: by training a set of T2I models from scratch (300M to 3B parameters), we find that larger models trained on more image data produce more accurate depth. Our strongest model is competitive with state-of-the-art monocular depth estimators and reduces AbsRel by 57% relative to existing joint image-depth generative models. These results provide strong evidence that image generation is a scalable pre-training objective for spatial perception.

Space configuration

Weights

All weights are pulled from public repos — no token required: bartduis/modality_forcing (DiT + FLUX.2 autoencoder, CC BY-NC 4.0) and Qwen/Qwen3-8B (text encoder). Override the model repo via the WEIGHTS_REPO Space variable.

Hardware

Pinned to zero-h200. The runner builds the DiT on the meta device and assigns checkpoint weights straight onto the GPU (runner.from_pretrained), so the BF16 model (~24 GB) loads in ~8 s on the first invocation — skipping ~45 s of throwaway random initialization — and stays resident for subsequent calls within the same Space instance.

Avoiding the cold re-download

A Space instance that has gone to sleep starts a fresh container, which re-downloads the weights (~24 GB DiT + ~16 GB Qwen3-8B + the FLUX.2 VAE) before the load above can run. To skip that, attach persistent storage (Settings → Storage, or mount an HF Bucket) and point the HF cache at it via the HF_HOME Space variable so the weights survive restarts.

License

Code: Apache-2.0 (files derived from the FLUX.2 reference implementation — flux_rgbd/_flux2/, flux_rgbd/dit.py — are Apache-2.0, Copyright Black Forest Labs, with World Labs modifications). Model weights: CC BY-NC 4.0.

Citation

If you find Modality Forcing useful, please consider citing:

@article{duisterhof2026mofo,
  title   = {Modality Forcing for Scalable Spatial Generation},
  author  = {Duisterhof, Bardienus Pieter and Ramanan, Deva and Ichnowski, Jeffrey and Johnson, Justin and Park, Keunhong},
  journal = {arXiv preprint arXiv:2606.13676},
  year    = {2026}
}