MODUS Delivers Unified Any-To-Any Multimodal Generation In Single Decoder

Chaining: generation can run in two modes. Chained: each modality conditions on the ones before it, e.g., Text → Depth → RGB. All outputs stay consistent with each other. Independent: each target is produced straight from the input in parallel, so they're free to vary. Chaining basically means the model’s output can be fed back in as input, so it can reach a final target by first generating a useful intermediate modality. This enables using a chain of intermediate modalities from a single model to address modality mappings that are harder to learn directly. As an example, we probe surface-normal estimation, comparing the direct RGB → Normal against routing through three candidate intermediates. Edge maps turn out to give the largest gain: they add pixel-aligned low-level geometry that complements surface-orientation estimation.