#01Oct 21, 2026
cs.CV
How Do DiTs Denoise? An Aggregator-Predictor View of Attention and MLPs
Gyeonghun Kang, William Xingxu Chen, Xiang Cheng
How do diffusion transformers (DiTs) represent image content and learn rules for reusing it to generalize? We develop an aggregator--predictor view of DiTs as input-adaptive local denoisers, in which context aggregation and patch prediction provide functional roles for interpreting attention and MLPs. We first establish that, for Gaussian Markov random field (GMRF) priors, the global Bayes denoiser can be approximated using only a subset of context patches, with error decaying exponentially in context radius. We further show that patchwise linear denoisers, often invoked to explain diffusion generalization, recover the empirical local denoiser with sufficiently informative aggregation. Within this view, AdaLN-modulated attention naturally parameterizes an input- and noise-dependent aggregator, while the MLP admits a learned key--value representation for tokenwise patch prediction. Our experiments support this view. DiTs trained on real images exhibit noise-dependent denoising contexts resembling those in the analytic GMRF setting, while attention mass exhibits a qualitatively corresponding broadening pattern. In a controlled CelebA64 study, shared-memory patch predictors used in place of DiT MLPs achieve competitive denoising and generative performance. Finally, cross-dataset component transplantation reveals compatibility patterns consistent with the proposed roles of QK in spatial weighting, VO in feature transport, and MLPs in patch prediction. Together, these results provide a component-level account of DiT denoising and motivate architecture-specific fine-tuning and controlled intervention.
#02Oct 21, 2026
cs.CV
JointLight: Human-Centric Video Relighting with Asynchronous RGB-Irradiance Joint Modeling
Henglei Lv, Bailin Deng, Xiaoqiang Liu and 4 more
Understanding and modeling light transport in human-centric scenes is a longstanding challenge. While recent progress in diffusion models has boosted research in relighting, precise illumination control still lags, primarily due to the scarcity of diverse, high-quality, precisely labeled video relighting datasets. Existing end-to-end relighting approaches model light transport implicitly through neural networks, which substantially complicates controllability. In contrast, some works resort to path-traced irradiance cues as an explicit blueprint for light transport. However, such constructed irradiance cues are prone to artifacts and hinder generative quality when directly used as control signals. With the key observation that noise corruption removes high-frequency artifacts and preserves low-frequency light transport, we introduce an asynchronous joint-modeling paradigm that facilitates fine-grained illumination control. During two-stage training, we force the model to jointly denoise RGB and ground-truth irradiance frames with separate timesteps scheduling. A domain matching strategy is applied to the second stage to bridge the domain gap between irradiance cues and ground-truth irradiance maps. At inference time, we introduce an asynchronous hijacked generation pipeline that facilitates flexible illumination control. For training data, we develop a multi-illumination high-quality synthetic video dataset with paired irradiance maps using a powerful game engine. Trained exclusively on this dataset, our model exhibits robust generalization to real- world scenarios and outperforms prior approaches in both lighting fidelity and generative quality. The code and dataset will be publicly released.
#03Oct 21, 2026
cs.CV
Expandable Modality Experts for Medical Vision-Language Learning
DM BAPPY, Eunsu Kim, Hanjoo Lee and 2 more
Learning across medical imaging modalities requires making use of images without paired text and accommodating new domains while retaining existing representations. We propose a medical vision-language technique built around an expandable bank of modality experts. Its Multiway Transformer combines shared visual computation with deterministically routed feed-forward experts and a separate text pathway. A modality-aware visual tokenizer supplies targets for masked pretraining on 19.4 million unpaired image rows across seven modalities; contrastive alignment then uses 3.23 million image-text pairs. Additional fundus and optical coherence tomography experts learn from paired data while the original image and text embedding paths remain frozen. With a ViT-Large backbone, the proposed technique achieves the best or tied-best result in 24 of 25 benchmark settings across nine modalities: 10 of 11 zero-shot classification benchmarks, all 10 linear probes, and all four retrieval benchmarks. Mean balanced accuracy is 0.674 for zero-shot classification and 0.856 for linear probing, and mean bidirectional retrieval recall is 20.70%. A separate three-seed sequential experiment with matched adaptation data yields fundus and OCT retrieval gains of 14.95 and 9.31 percentage points, with no measured loss on preceding fundus or pathology tasks. The preservation argument requires unchanged existing computation paths and inference conditions, and feature-identity checks verify retention on evaluated inputs. The proposed technique provides a modular approach to learning and expanding medical vision-language representations.
#04Oct 21, 2026
cs.CV
Attn-Forcing: Learning Full-Context Attention Proxies for Streaming Video Generation
Junyao Gao, Sibo Liu, Jiaxing Li and 5 more
Increasing the attention context length significantly improves streaming video generation quality, while incurring prohibitive memory and computational costs as the context grows. Existing methods attempt to capture this contextual information through attention selection or compression, but remain limited by bounded representations and lack explicit attention-level supervision to retain historical contributions to current queries. To address this, we introduce Attn-Forcing, a novel framework that preserves full-context attention contributions for streaming video generation under bounded memory and computation. We achieve this by introducing attention proxies that continuously approximate the attention responses induced by the historical context. Specifically, we implement the proxies as learnable visual latents and employ a Proxy Update Module to incorporate visual context from historical latents. Moreover, we propose an Attention Matching Loss to explicitly constrain the attention behavior of the learned proxies using full-context attention supervision. Extensive qualitative and quantitative experiments demonstrate that Attn-Forcing consistently improves streaming video generation quality and achieves state-of-the-art performance over existing baseline methods.
#05Oct 21, 2026
cs.CV
Asking a Multimodal LLM Whether Someone Fell: Zero-Shot Screening of Low-Resolution Thermal Video in a Laboratory and an Occupied Home
Marcos Junca, Mauro Junca
Thermal cameras can watch over older people living alone without capturing faces, and multimodal large language models (LLMs) can judge a short video without task-specific training. We evaluate a production multimodal LLM, Gemini 3.8 Flash at temperature 0, as the decision stage of an event-triggered fall screener, used exactly as a deployed prototype uses it: one request per short thermal clip, with its sound track when there is one, and its predecessor Gemini 3.7 Flash on the same requests. On 438 staged falls and 563 non-fall windows that we located by hand in a public 32x24 thermal dataset, the deployed request called 69% of the held-out subjects' falls (67% over all subjects; bed falls excluded) and none of 499 non-fall windows, including people lying on the floor and getting up. Over 76 h of ordinary life in an occupied room, recorded at 160x120 with sound, it called none of 395 clips a fall (95% upper bound 1.2 false alarms per day) and caught 4 of the 9 staged falls the camera could see. One cautionary sentence in the prompt ("a person who goes out of view ... is not a fall") accounted for many of the misses: removing it raised sensitivity on the held-out subjects from 69% to 83% on 3.8 Flash and from 80% to 92% on Gemini 3.7 Flash, and in the home from 4 to 9 of the 9 visible falls, with no false call in ordinary living but at the price of calling deliberate lie-downs a fall. Sensitivity collapsed at 8x6 pixels (34%); identical requests disagreed on 4% of windows; and when the model called a fall, its explanation placed it within a second of the annotated landing 99% of the time. Two conventional detectors trained on the laboratory data caught nearly every held-out laboratory fall but raised 14.2 and 23.9 false alarms per day in the home. We release the frame-level fall annotations of the public dataset and every model answer on it.