MMDiff: Extending Diffusion Transformers for Multi-Modal Generation

Text-to-image diffusion transformers learn about objects and scenes by learning to generate them, making them strong candidates for training-free zero-shot open-vocabulary semantic segmentation. State-of-the-art attribution methods score each pixel independently, comparing its features against a fixed text-derived class representation, whether as an output-space similarity or as a cross-attention weight. This discards structured signals the model itself exposes: the temporal structure of the generative trajectory, the visual appearance statistics of each concept, and the image's own pairwise feature geometry. We present MAVISEG, a training-free refinement layer that recovers these signals. Because its operators consume only a pixel-by-concept score field and a pixel feature space, MAVISEG is capture-agnostic rather than tied to one attribution method. Across six benchmarks it achieves the strongest overall results among training-free methods, including the best mIoU on every benchmark. Interestingly, gains are largest where the initial capture is weakest, and individual operators contribute depending on the noise in the field they refine. Our results indicate that diffusion transformers carry more concept-level information than current attribution methods recover, and that much of it is lost on the way to the mask rather than absent from the model.

Diffusion transformers (DiTs) have emerged as a dominant architecture for text-to-image generation, yet their performance drops when generating at resolutions beyond their training range. Existing training-free approaches mitigate this by modifying inference-time attention behavior, often through Rotary Position Embeddings (RoPE) extrapolation combined with attention scaling. However, these strategies apply a uniform and content-agnostic scaling across RoPE components with distinct frequency characteristics, inducing a trade-off between preserving global structure and recovering fine detail. We introduce SEGA, a training-free method that dynamically scales attention across RoPE components according to the latent's spatial-frequency structure at each denoising step. This adaptive scaling improves both structural coherence and fine-detail fidelity. Experiments show that SEGA consistently improves high-resolution synthesis across multiple target resolutions, outperforming state-of-the-art training-free baselines.

Diffusion models are primarily trained for image synthesis, yet their denoising trajectories encode rich, spatially aligned visual priors. In this paper, we demonstrate that these priors can be utilized for text-conditioned semantic and open-vocabulary segmentation, and this approach can be generalized to various downstream tasks to make a general-purpose diffusion segmentation framework. Concretely, we introduce DiGSeg (Diffusion Models as a Generalist Segmentation Learner), which repurposes a pretrained diffusion model into a unified segmentation framework. Our approach encodes the input image and ground-truth mask into the latent space and concatenates them as conditioning signals for the diffusion U-Net. A parallel CLIP-aligned text pathway injects language features across multiple scales, enabling the model to align textual queries with evolving visual representations. This design transforms an off-the-shelf diffusion backbone into a universal interface that produces structured segmentation masks conditioned on both appearance and arbitrary text prompts. Extensive experiments demonstrate state-of-the-art performance on standard semantic segmentation benchmarks, as well as strong open-vocabulary generalization and cross-domain transfer to medical, remote sensing, and agricultural scenarios-without domain-specific architectural customization. These results indicate that modern diffusion backbones can serve as generalist segmentation learners rather than pure generators, narrowing the gap between visual generation and visual understanding.