#01Nov 19, 2026
cs.CV
RL Post-Training in Diffusion Transformers Is Activation Steering
Tuna Han Salih Meral, Kaan Oktay, Alec Helbling and 3 more
Reinforcement learning (RL) post-training substantially improves the visual quality, text rendering, and compositional accuracy of diffusion transformers. Because capabilities such as rendering requested text or composing multiple objects are inherently prompt-dependent, a natural hypothesis is that improving them requires learning complex, input-dependent transformations. To isolate how much of these gains an input-independent shift explains, we probe each RL-tuned model with a deliberately simple parameterization: a fixed set of steering vectors added to the block outputs of the frozen base model. On SD3.5-M, these vectors hold 65x fewer parameters than the RL-tuned LoRA (0.29M against 18.8M). We fit them to match the velocity predictions of public RL-tuned checkpoints, with no reward in the objective. Although the vectors are identical for every prompt, token, and sampling seed, they recover 90.0-96.3% of the reward gains of six RL-tuned image models, and 89.8% and 111.9% on two RL-tuned video models. Initialized at zero and trained with online RL, the vectors alone exceed the reward gains of weight-based RL on text rendering (OCR) and human preference (PickScore) on SD3.5-M. A constant shift thus suffices for most of the gain of policy-gradient RL post-training: without new input-dependent transformations, it steers the pretrained generator toward what it can already compute.
#02Nov 19, 2026
cs.CV
AURA: AUdio-dRiven streaming Avatar
Nathaniel Cohen, Nicolas Dufour, Amélie Royer and 2 more
Current audio-driven video diffusion models can synthesize realistic talking-head videos but generally require dozens of diffusion denoising steps over multi-second video chunks, making them unsuitable for streaming applications such as voice assistants and virtual avatars, where low-latency and online causal generation are required properties. AURA generates video chunk by chunk with a rolling KV cache, requiring only the audio of the current chunk to start generating, rather than a multi-second audio look-ahead like bidirectional models. It comes in two sizes: AURA 1.3B streams 576p video in real time on a single GPU (20.7 fps for a 16 Hz playback rate), making it suited to interactive avatar applications, and AURA 5B streams 720p video at 9.7 fps. Our two-stage distillation pipeline first aligns the causal student with the teacher, by training it to match the teacher's trajectories with fewer diffusion steps and a chunk-wise sliding window causal attention mask, then applies Distribution Matching Distillation (DMD) with backward simulation to close the remaining quality gap. On standard talking-head benchmarks, AURA 1.3B runs 23x faster than its own bidirectional teacher and reaches a quality comparable to TalkVerse at a lower resolution, and AURA 5B matches its TalkVerse teacher at 720p while running 17x faster. To support reproducibility and further research on streaming avatar generation, we release the weights of both models along with a dataset of 30K synthetic videos and their precomputed teacher trajectories.
#03Nov 19, 2026
cs.CV
FFGaussType: Feed-Forward Conversion of Glyph and SVG Geometry into Compact Gaussian Assets
Haoan Feng, Tianfu Wang, Peimeng Guan and 2 more
Immersive 3D content allows text and graphical elements, such as subtitles, annotations, and icons, to be placed inside a scene rather than displayed as flat overlays. Placed this way, they respect occlusion and perspective and take on user-assigned materials and lighting while preserving their original shapes. In Gaussian scenes, however, each new glyph or icon requires its own iterative optimization, which limits scalable content creation. We present FFGaussType, which converts extruded glyph and SVG outlines into Gaussian assets in a single forward pass. A point encoder maps anchors sampled on the known surface to a single feature plane, from which a decoder shared by all assets predicts neural Gaussians carrying surface normals. This separates geometric conversion from material and lighting choices, and the feature plane supports image and video coding for compact delivery. We construct a paired dataset of 27,221 glyph and SVG icon assets with anchors and calibrated normal and silhouette renderings, holding out font families for evaluation. Conversion from prepared anchors takes 9.1 ms per asset, at a fraction of the compute and memory of prior baselines. On unseen font families and test icons, FFGaussType achieves $1.62^\circ$ normal error and $0.985$ silhouette IoU, the best among the compared feed-forward converters, matching the fidelity that per-asset fitting needs thousands of steps to reach. The code and dataset will be released after publication.
#04Nov 19, 2026
cs.CV
SGAR: Feed-Forward 3D Reconstruction via Sparse Geometry-Guided Attention
Aoxiang Fan, Maria Pilligua, Corentin Dumery and 2 more
Feed-forward models have been very successful in 3D reconstruction with ad hoc attention designs that relate all view patches densely. However, the cost of this dense global attention grows quadratically with the number of views, preventing the model from applying to larger-scale reconstruction settings. In this paper, we argue that most dense attention computations are unnecessary from a classical structure-from-motion perspective: 3D structures stem from geometrically matching regions in 2D. We bring this principle to learning-based feed-forward reconstruction with our Sparse Geometric Attention Reconstruction (SGAR) method. In SGAR, each patch token attends to only a small set of patch tokens from other views, chosen by a lightweight learned key selection module trained to predict cross-view 3D proximity of tokens. Thus, dense global attention is replaced by much more efficient geometric-similarity-guided sparse attention. SGAR is a general method that can be applied to different feed-forward models by fine-tuning. In our experiments, we show that SGAR is the fastest among both state-of-the-art dense and sparse methods and the most accurate sparse one. It runs up to 12-times faster than dense attention at 1,000 views with stable pose accuracy where existing sparse methods fail, and shows the best performance in reconstructing more complete geometries on the same time budget.
#05Nov 19, 2026
cs.CV
ScaleWorld: High-Resolution World Generation through Shared Prediction and Refinement
Yansong Qu, Wesker Zheng, Quanjian Song and 4 more
High-resolution generation enables video world models to render fine-grained scene structures and textures, but directly generating high-resolution video incurs substantial computational and memory costs. We present ScaleWorld, a causal video model that unifies low-resolution world prediction and high-resolution refinement within a single shared backbone. By refining its own predictions, ScaleWorld produces high-resolution world videos without requiring a separate refinement model. For each video chunk, the shared backbone first completes a low-resolution prediction and then performs one-step high-resolution refinement conditioned on that prediction. To support efficient streaming generation, we maintain separate causal states with a long low-resolution history for prediction and a short high-resolution history for refinement. To address task interference during distillation, we separate specialist optimization from capability consolidation and introduce task-asymmetric on-policy distillation (TA-OPD). This procedure transfers prediction and refinement behaviors from specialized teachers into a shared student through local velocity matching under their respective sampling and history policies. {Extensive experiments demonstrate that ScaleWorld achieves high-quality world generation with improved computational and memory efficiency.