DyPE: Dynamic Position Extrapolation for Ultra High Resolution Diffusion

Diffusion Transformer models can generate images with remarkable fidelity and detail, yet training them at ultra-high resolutions remains extremely costly due to the self-attention mechanism's quadratic scaling with the number of image tokens. In this paper, we introduce Dynamic Position Extrapolation (DyPE), a novel, training-free method that enables pre-trained diffusion transformers to synthesize images at resolutions far beyond their training data, with no additional sampling cost. DyPE takes advantage of the spectral progression inherent to the diffusion process, where low-frequency structures converge early, while high-frequencies take more steps to resolve. Specifically, DyPE dynamically adjusts the model's positional encoding at each diffusion step, matching their frequency spectrum with the current stage of the generative process. This approach allows us to generate images at resolutions that exceed the training resolution dramatically, e.g., 16 million pixels using FLUX. On multiple benchmarks, DyPE consistently improves performance and achieves state-of-the-art fidelity in ultra-high-resolution image generation, with gains becoming even more pronounced at higher resolutions. Project page is available at https://noamissachar.github.io/DyPE/.

Let it Snow! Animating Static Gaussian Scenes With Dynamic Weather Effects

3D Gaussian Splatting has recently enabled fast and photorealistic reconstruction of static 3D scenes. However, introducing dynamic elements that interact naturally with such static scenes remains challenging. Accordingly, we present a novel hybrid framework that combines Gaussian-particle representations for incorporating physically-based global weather effects into static 3D Gaussian Splatting scenes, correctly handling the interactions of dynamic elements with the static scene. We follow a three-stage process: we first map static 3D Gaussians to a particle-based representation. We then introduce dynamic particles and simulate their motion using the Material Point Method (MPM). Finally, we map the simulated particles back to the Gaussian domain while introducing appearance parameters tailored for specific effects. To correctly handle the interactions of dynamic elements with the static scene, we introduce specialized collision handling techniques. Our approach supports a variety of weather effects, including snowfall, rainfall, fog, and sandstorms, and can also support falling objects, all with physically plausible motion and appearance. Experiments demonstrate that our method significantly outperforms existing approaches in both visual quality and physical realism.

RewardSDS: Aligning Score Distillation via Reward-Weighted Sampling

Score Distillation Sampling (SDS) has emerged as an effective technique for leveraging 2D diffusion priors for tasks such as text-to-3D generation. While powerful, SDS struggles with achieving fine-grained alignment to user intent. To overcome this, we introduce RewardSDS, a novel approach that weights noise samples based on alignment scores from a reward model, producing a weighted SDS loss. This loss prioritizes gradients from noise samples that yield aligned high-reward output. Our approach is broadly applicable and can extend SDS-based methods. In particular, we demonstrate its applicability to Variational Score Distillation (VSD) by introducing RewardVSD. We evaluate RewardSDS and RewardVSD on text-to-image, 2D editing, and text-to-3D generation tasks, showing significant improvements over SDS and VSD on a diverse set of metrics measuring generation quality and alignment to desired reward models, enabling state-of-the-art performance. Project page is available at https://itaychachy.github.io/reward-sds/.

Structurally Disentangled Feature Fields Distillation for 3D Understanding and Editing

Recent work has demonstrated the ability to leverage or distill pre-trained 2D features obtained using large pre-trained 2D models into 3D features, enabling impressive 3D editing and understanding capabilities using only 2D supervision. Although impressive, models assume that 3D features are captured using a single feature field and often make a simplifying assumption that features are view-independent. In this work, we propose instead to capture 3D features using multiple disentangled feature fields that capture different structural components of 3D features involving view-dependent and view-independent components, which can be learned from 2D feature supervision only. Subsequently, each element can be controlled in isolation, enabling semantic and structural understanding and editing capabilities. For instance, using a user click, one can segment 3D features corresponding to a given object and then segment, edit, or remove their view-dependent (reflective) properties. We evaluate our approach on the task of 3D segmentation and demonstrate a set of novel understanding and editing tasks.

Designing a Conditional Prior Distribution for Flow-Based Generative Models

Flow-based generative models have recently shown impressive performance for conditional generation tasks, such as text-to-image generation. However, current methods transform a general unimodal noise distribution to a specific mode of the target data distribution. As such, every point in the initial source distribution can be mapped to every point in the target distribution, resulting in long average paths. To this end, in this work, we tap into a non-utilized property of conditional flow-based models: the ability to design a non-trivial prior distribution. Given an input condition, such as a text prompt, we first map it to a point lying in data space, representing an ``average" data point with the minimal average distance to all data points of the same conditional mode (e.g., class). We then utilize the flow matching formulation to map samples from a parametric distribution centered around this point to the conditional target distribution. Experimentally, our method significantly improves training times and generation efficiency (FID, KID and CLIP alignment scores) compared to baselines, producing high quality samples using fewer sampling steps.

Through-The-Mask: Mask-based Motion Trajectories for Image-to-Video Generation

We consider the task of Image-to-Video (I2V) generation, which involves transforming static images into realistic video sequences based on a textual description. While recent advancements produce photorealistic outputs, they frequently struggle to create videos with accurate and consistent object motion, especially in multi-object scenarios. To address these limitations, we propose a two-stage compositional framework that decomposes I2V generation into: (i) An explicit intermediate representation generation stage, followed by (ii) A video generation stage that is conditioned on this representation. Our key innovation is the introduction of a mask-based motion trajectory as an intermediate representation, that captures both semantic object information and motion, enabling an expressive but compact representation of motion and semantics. To incorporate the learned representation in the second stage, we utilize object-level attention objectives. Specifically, we consider a spatial, per-object, masked-cross attention objective, integrating object-specific prompts into corresponding latent space regions and a masked spatio-temporal self-attention objective, ensuring frame-to-frame consistency for each object. We evaluate our method on challenging benchmarks with multi-object and high-motion scenarios and empirically demonstrate that the proposed method achieves state-of-the-art results in temporal coherence, motion realism, and text-prompt faithfulness. Additionally, we introduce \benchmark, a new challenging benchmark for single-object and multi-object I2V generation, and demonstrate our method's superiority on this benchmark. Project page is available at https://guyyariv.github.io/TTM/.

Intermediate Representations for Enhanced Text-To-Image Generation Using Diffusion Models

Text-to-image diffusion models have demonstrated an impressive ability to produce high-quality outputs. However, they often struggle to accurately follow fine-grained spatial information in an input text. To this end, we propose a compositional approach for text-to-image generation based on two stages. In the first stage, we design a diffusion-based generative model to produce one or more aligned intermediate representations (such as depth or segmentation maps) conditioned on text. In the second stage, we map these representations, together with the text, to the final output image using a separate diffusion-based generative model. Our findings indicate that such compositional approach can improve image generation, resulting in a notable improvement in FID score and a comparable CLIP score, when compared to the standard non-compositional baseline.

Improving Visual Commonsense in Language Models via Multiple Image Generation

Commonsense reasoning is fundamentally based on multimodal knowledge. However, existing large language models (LLMs) are primarily trained using textual data only, limiting their ability to incorporate essential visual information. In contrast, Visual Language Models, which excel at visually-oriented tasks, often fail at non-visual tasks such as basic commonsense reasoning. This divergence highlights a critical challenge - the integration of robust visual understanding with foundational text-based language reasoning. To this end, we introduce a method aimed at enhancing LLMs' visual commonsense. Specifically, our method generates multiple images based on the input text prompt and integrates these into the model's decision-making process by mixing their prediction probabilities. To facilitate multimodal grounded language modeling, we employ a late-fusion layer that combines the projected visual features with the output of a pre-trained LLM conditioned on text only. This late-fusion layer enables predictions based on comprehensive image-text knowledge as well as text only when this is required. We evaluate our approach using several visual commonsense reasoning tasks together with traditional NLP tasks, including common sense reasoning and reading comprehension. Our experimental results demonstrate significant superiority over existing baselines. When applied to recent state-of-the-art LLMs (e.g., Llama3), we observe improvements not only in visual common sense but also in traditional NLP benchmarks. Code and models are available under https://github.com/guyyariv/vLMIG.

Coarse-To-Fine Tensor Trains for Compact Visual Representations

The ability to learn compact, high-quality, and easy-to-optimize representations for visual data is paramount to many applications such as novel view synthesis and 3D reconstruction. Recent work has shown substantial success in using tensor networks to design such compact and high-quality representations. However, the ability to optimize tensor-based representations, and in particular, the highly compact tensor train representation, is still lacking. This has prevented practitioners from deploying the full potential of tensor networks for visual data. To this end, we propose 'Prolongation Upsampling Tensor Train (PuTT)', a novel method for learning tensor train representations in a coarse-to-fine manner. Our method involves the prolonging or `upsampling' of a learned tensor train representation, creating a sequence of 'coarse-to-fine' tensor trains that are incrementally refined. We evaluate our representation along three axes: (1). compression, (2). denoising capability, and (3). image completion capability. To assess these axes, we consider the tasks of image fitting, 3D fitting, and novel view synthesis, where our method shows an improved performance compared to state-of-the-art tensor-based methods. For full results see our project webpage: https://sebulo.github.io/PuTT_website/

DGD: Dynamic 3D Gaussians Distillation

We tackle the task of learning dynamic 3D semantic radiance fields given a single monocular video as input. Our learned semantic radiance field captures per-point semantics as well as color and geometric properties for a dynamic 3D scene, enabling the generation of novel views and their corresponding semantics. This enables the segmentation and tracking of a diverse set of 3D semantic entities, specified using a simple and intuitive interface that includes a user click or a text prompt. To this end, we present DGD, a unified 3D representation for both the appearance and semantics of a dynamic 3D scene, building upon the recently proposed dynamic 3D Gaussians representation. Our representation is optimized over time with both color and semantic information. Key to our method is the joint optimization of the appearance and semantic attributes, which jointly affect the geometric properties of the scene. We evaluate our approach in its ability to enable dense semantic 3D object tracking and demonstrate high-quality results that are fast to render, for a diverse set of scenes. Our project webpage is available on https://isaaclabe.github.io/DGD-Website/