# Diffusion Transformer

MAAS replaces the text-embedding conditioning of a pretrained audio diffusion transformer with region-of-interest fMRI responses, removing the decoder that previously predicted an external stimulus representation such as CLAP or YamNet. On a 7T fMRI dataset it improves acoustic and semantic metrics, including CLAP similarity, pairwise identification accuracy and Mel-SSIM, over baselines such as two-stage fMRI-to-latent conditioning. The framework also uses a fidelity metric grounded in a computational model of auditory cortex to quantify how individual auditory regions contribute to the reconstructed audio.

MAAS replaces the text-embedding conditioning of a pretrained audio diffusion transformer with region-of-interest fMRI responses, removing the decoder that previously predicted an external stimulus representation such as CLAP or YamNet. On a 7T fMRI dataset it improves acoustic and semantic metrics, including CLAP similarity, pairwise identification accuracy and Mel-SSIM, over baselines such as two-stage fMRI-to-latent conditioning. The framework also uses a fidelity metric grounded in a computational model of auditory cortex to quantify how individual auditory regions contribute to the reconstructed audio.

A training-free diffusion caching rule spends its exact denoiser evaluations where they matter by dividing shallow-probe output change by input displacement, lowering median LPIPS at matched exact-compute budgets on FLUX.1-dev, Wan2.1 and Open-Sora 1.2.

A training-free diffusion caching rule spends its exact denoiser evaluations where they matter by dividing shallow-probe output change by input displacement, lowering median LPIPS at matched exact-compute budgets on FLUX.1-dev, Wan2.1 and Open-Sora 1.2.

A controlled from-scratch comparison shows that adding long skip connections to a DiT improves its intermediate features across semantic correspondence, segmentation and depth, and the gain survives even when generation quality is matched by FID.

A controlled from-scratch comparison shows that adding long skip connections to a DiT improves its intermediate features across semantic correspondence, segmentation and depth, and the gain survives even when generation quality is matched by FID.

Across the text-to-image Transformers FLUX.1, Z-Image and Infinity-2B, attention sinks land in architecture-specific places: the autoregressive model concentrates on its first global token, while the bidirectional diffusion models concentrate near modality boundaries, inside text regions and at image corners.

Across the text-to-image Transformers FLUX.1, Z-Image and Infinity-2B, attention sinks land in architecture-specific places: the autoregressive model concentrates on its first global token, while the bidirectional diffusion models concentrate near modality boundaries, inside text regions and at image corners.

The authors remove the residual connections around Self-Attention and MLP in the condition encoder of the Decoupled Diffusion Transformer (DDT), stabilize training with zero-initialized output gates and a mixed orthogonal attention initialization, and form the encoder output by aggregating the input patch embedding with the outputs of all encoder blocks; under ImageNet-1k training at the stated resolution, patch size 2, 80 epochs and batch size 1024, DDT-RFE improves overall over DDT on image classification, semantic segmentation, object discovery and semantic correspondence while using fewer encoder blocks (12 vs. 20 at L, 14 vs. 22 at XL), and lowers ImageNet generation FID from 8.98 to 7.68 at L and from 7.27 to 6.81 at XL.

The authors remove the residual connections around Self-Attention and MLP in the condition encoder of the Decoupled Diffusion Transformer (DDT), stabilize training with zero-initialized output gates and a mixed orthogonal attention initialization, and form the encoder output by aggregating the input patch embedding with the outputs of all encoder blocks; under ImageNet-1k training at the stated resolution, patch size 2, 80 epochs and batch size 1024, DDT-RFE improves overall over DDT on image classification, semantic segmentation, object discovery and semantic correspondence while using fewer encoder blocks (12 vs. 20 at L, 14 vs. 22 at XL), and lowers ImageNet generation FID from 8.98 to 7.68 at L and from 7.27 to 6.81 at XL.

The work proposes FADE, a multi-concept unlearning framework for text-to-video diffusion transformers that first applies a joint closed-form key/value edit suppressing all target concepts, then trains per-concept frame-aware low-rank adapters gated by the frame index and denoising timestep to remove residual per-frame leakage, and combines adapters by a similarity-based soft router; erasing 16 concepts (objects, artistic styles, and nudity) from a single Wan2.1-T2V-1.3B backbone, FADE reduces residual object-benchmark accuracy to 4.9% against 15.5% for the strongest of eight baselines, keeps the VBench average within 0.9% of the unedited model, and the ranking holds under a VLM judge and a blinded human study.

The work proposes FADE, a multi-concept unlearning framework for text-to-video diffusion transformers that first applies a joint closed-form key/value edit suppressing all target concepts, then trains per-concept frame-aware low-rank adapters gated by the frame index and denoising timestep to remove residual per-frame leakage, and combines adapters by a similarity-based soft router; erasing 16 concepts (objects, artistic styles, and nudity) from a single Wan2.1-T2V-1.3B backbone, FADE reduces residual object-benchmark accuracy to 4.9% against 15.5% for the strongest of eight baselines, keeps the VBench average within 0.9% of the unedited model, and the ranking holds under a VLM judge and a blinded human study.