Prompt2Effect synthesizes effect-specific LoRA weights for image-to-video diffusion models in a single forward pass, reducing per-effect specialization from 56 GPU hours to 3.3 seconds.

Canvas-to-Image introduces a unified framework that consolidates heterogeneous controls (subject references, bounding boxes, pose skeletons) into a single canvas interface for high-fidelity compositional image generation.

Diffusion-DRF uses a frozen VLM critic to provide free, rich, and differentiable feedback for stable video diffusion fine-tuning.

EasyV2V: A high-quality instruction-based video editing framework that enables intuitive video manipulation through natural language instructions.

Omni-Attribute can isolate a specific attribute, whether it is an abstract concept or not, from any image and merge those selected attributes from multiple images into a coherent generation.

We prevent shortcuts in adapter training by explicitly providing the shortcuts during training, forcing the model to learn more robust representations.

ComposeMe is a human-centric generative model that enables disentangled control over multiple visual attributes — such as identity, hair, and garment — across multiple subjects, while also supporting text-based control.

ThinkDiff enables multimodal in-context reasoning in diffusion models by aligning vision-language models to LLM decoders, transferring reasoning capabilities without requiring complex reasoning-based datasets.

Omni-ID is a novel facial representation tailored for generative tasks, encoding identity features from unstructured images into a fixed-size representation that captures diverse expressions and poses.

WonderLand is a video-latent based approach for single-image 3D reconstruction in large-scale scenes.

AC3D studies when and how you should condition camera signals into a video diffusion model for a better camera control and a higher video quality.

Magic123 proposes a hybrid score distillation algorithm and a coarse-to-fine image-to-3D pipeline that produces high-quality high-resolution 3D content from a single unposed image.

Pix4Point shows that image pretraining significantly improves point cloud understanding.

PointNeXt boosts the performance of PointNet++ to the state-of-the-art level with improved training and scaling strategies.

ASSANet makes PointNet++ faster and more accurate.

This work transfers concepts such as residual/dense connections and dilated convolutions from CNNs to GCNs in order to successfully train very deep GCNs.