The paper revisits the interpretation of depth profiles in Diffusion Transformer backbones, showing that the commonly observed inverted‑U curve reflects the influence of the prediction target (e.g., ε, u, x0) rather than a simple shift from coarse to fine semantic granularity. It argues that the semantically meaningful signal is merely an intermediate by‑product of predicting the target, and that different training setups—particularly the choice of prediction target—predictably alter the curve’s shape. The authors demonstrate how this new understanding can guide the optimal placement of REPA modules and provide code for further exploration.
arXiv:2605. 20708v2 Announce Type: replace-cross Abstract: Diffusion Transformers (DiTs) have become a de facto backbone of modern visual generation, and nearly every major axis of their design -- tokenization, attention, conditioning, objectives, and latent autoencoders -- has been extensively revisited.
The study investigates whether transformer models transport token representations across layers in a manner consistent with optimal transport theory. Using Pythia-160m and Pythia-410m, the authors measure exact assignments between consecutive layer clouds, calibrate couplings, and decompose the transport cost into common shifts and token‑specific moves. Results show that at the final layer both models align closely with the optimal‑transport map—Pythia‑410m at optimal cost and Pythia‑160m slightly above—while earlier layers deviate, and the agreement improves with training.
arXiv:2606. 26538v1 Announce Type: cross Abstract: Deep Transformers are composed of uniformly stacked residual blocks, yet their deepest layers often add little value.
arXiv:2602. 13298v4 Announce Type: replace-cross Abstract: This paper presents a controlled comparative study of convolutional neural network (CNN) topology and image classification performance across the architectural families VGG, ResNet, and GoogLeNet, evaluated on CIFAR-10 under a unified training protocol.
The paper investigates why Diffusion Transformers (DiT) produce weaker features for vision tasks compared to their strong generative performance. Through controlled training, the authors demonstrate that incorporating long skip connections (LSCs) significantly improves feature quality. They explain that LSCs allow diffusion noise to bypass the main backbone, preventing noise pollution of image signals and aligning the architecture more closely with the diffusion objective.
The paper investigates why Diffusion Transformers perform well when predicting clean data but poorly when predicting noise or velocity. It introduces the concept of *residual-stream burden*, explaining that noisy targets force the residual stream to carry noise-dependent variation through the network, limiting representation flexibility. Experiments show that spectral concentration of clean targets reduces this burden, and the authors propose Spatially Indexed Hyper-Connections (SiHC) to expand residual bandwidth, achieving state‑of‑the‑art FID on ImageNet.
arXiv:2607. 13491v1 Announce Type: cross Abstract: Looped Transformers scale sequential computation by applying a compact stack of physical blocks for multiple rounds, increasing unrolled depth without increasing stored parameters.
Recent advances in Diffusion Transformers (DiTs) have enabled remarkable progress in visual synthesis, benefiting from their superior scalability. To facilitate DiTs' capability of capturing meaningful internal representations, recent works such as REPA incorporate external pretrained encoders for representation alignment.
The paper investigates how one-step generative models, which compress multi-step diffusion processes, handle denoising across a single forward pass. It introduces the concept of "depth as time," showing that intermediate layers of a model can be decoded to reveal a denoising trajectory similar to multi-step diffusion. The study finds that models trained with a time-indexed transport task, such as MeanFlow, exhibit this depthwise denoising and can be highly compressed, whereas models without such a task do not show the same behavior.
We investigate how each component of the Transformer feedforward block architecture design determines how much rank survives across depth at initialization. We reinterpret skip connections and normalization, long understood as controlling magnitude, as mechanisms for preserving gradient rank across depth, since the very matrix multiplications and nonlinear activations that make the network expressive also reduce the rank.
arXiv:2609.07222v1 Announce Type: new Abstract: Transformer surrogates for 3D flow prediction compress an industrial mesh into a small set of tokens from which every prediction point reads. Two opera...
The paper introduces the first controlled benchmark for optimizers in discrete diffusion models, evaluating seven optimizers (AdamW, Lion, Muon, SOAP, MARS, MARS‑M, Schedule‑Free) across four diffusion formulations: masked diffusion on text8, uniform diffusion on QM9 and LM1B, and Gaussian diffusion on CelebA‑64. Each optimizer undergoes the same search protocol and is retrained with full budget and multiple seeds, revealing that AdamW, while strong, is not universally optimal and that optimizers validated on autoregressive language models (Muon, MARS‑M, SOAP) can outperform tuned AdamW on certain tasks.