Image Generation

Concept-Incremental Flexible Customization (CIFC) model tackles catastrophic forgetting and concept neglect in continually adapting text-to-image diffusion models, enabling flexible personalization.

Diffusion models surprisingly learn factorized representations, enabling compositional generalization, but struggle with interpolation; training with independent factors drastically improves data effi…

Hollowed Net efficiently personalizes text-to-image diffusion models on-device by temporarily removing deep U-Net layers during training, drastically reducing memory usage without sacrificing performa…

AICT: Adaptive-Interval Color Transformation harmonizes high-resolution images by predicting pixel-wise color changes, adaptively adjusting sampling intervals to capture local variations, and using a …

Visualizing high-dimensional posterior distributions is challenging. This paper introduces ‘Posterior Trees,’ a novel method using tree-structured neural network predictions for hierarchical uncertai…

HiCo: Hierarchical Controllable Diffusion Model achieves superior layout-to-image generation by disentangling spatial layouts through a multi-branch network structure, resulting in high-quality images…

HairFastGAN achieves realistic and robust hairstyle transfer in near real-time using a novel encoder-based approach, significantly outperforming optimization-based methods.

HairDiffusion uses latent diffusion models and a multi-stage blending technique to achieve vivid, multi-colored hair editing in images, preserving other facial features.

Boost image quality in diffusion models without reducing variation using Autoguidance: guide a high-quality model with a less-trained version of itself!

This paper introduces a novel gradient-free decoder inversion method for latent diffusion models, improving efficiency and memory usage compared to existing gradient-based methods. The method is theo…

This paper introduces a goal-conditioned reinforcement learning approach that efficiently tunes photo finishing pipelines, achieving high-quality results in fewer iterations than optimization-based me…

GenWarp generates high-quality novel image views from a single input image by using a semantic-preserving generative warping framework, outperforming existing methods.

This paper introduces a novel multi-stage generation framework for creating compositional scenes with fine-grained control by leveraging a trained diffusion model to produce individual scene component…

One-shot animatable head avatar reconstruction is achieved using a novel dual-lifting method that generates 3D Gaussians from a single image, enabling real-time expression control and rendering with s…

Part-Aware Diffusion Model (PA-Diffusion) enables precise and efficient manipulation of articulated objects in real images by using abstract 3D models and dynamic feature maps, overcoming limitations …

FuseAnyPart: Swap facial parts seamlessly using multiple reference images via diffusion, achieving high-fidelity results.

Diffusion models, while excelling in image generation, are vulnerable to data poisoning. This paper demonstrates a BadNets-like attack’s effectiveness against diffusion models, causing image misalign…

FreqMark: Robust invisible image watermarking via latent frequency space optimization, resisting regeneration attacks and achieving >90% bit accuracy with high image quality.

FeINFN: a novel Fourier-enhanced Implicit Neural Fusion Network, achieves state-of-the-art hyperspectral image fusion by innovatively combining spatial and frequency information in both the spatial an…

This work proposes FACL, a novel loss function for precipitation nowcasting, improving forecast sharpness and meteorological skill without sacrificing accuracy.