Image Generation

Self-Supervised Controllable Generation (SCG) framework achieves brain-like associative generation by using a modular autoencoder with equivariance constraints and a self-supervised pattern completion…

AI now draws almost as well as humans, thanks to novel latent diffusion model regularizations that mimic human cognitive biases.

A novel data-driven relighting model achieves state-of-the-art accuracy by learning latent intrinsic and extrinsic scene properties, even recovering albedo without explicit supervision.

KOALA: New efficient text-to-image diffusion models achieving 4x speed and 69% size reduction of SDXL, generating 1024px images on consumer GPUs with 8GB VRAM.

Pixelsmith: Generate gigapixel images with a single GPU, surpassing limitations of existing methods through a cascading approach and innovative guidance mechanism.

IR-CM: One-step image restoration using a novel consistency model for fast and universal performance.

IODA achieves efficient one-shot domain adaptation for super-resolution using a novel instance-guided strategy and image-level domain alignment, significantly improving performance with limited target…

Invertible Consistency Distillation (iCD) achieves high-quality image editing in ~7 steps by enabling both fast editing and strong generation using a generalized distillation framework and dynamic cla…

Researchers model a manifold of customized diffusion models as a subspace of weights, enabling controllable creation of new models via sampling, editing, and inversion from a single image.

Interpretable lightweight transformers are built by unrolling graph smoothness priors, achieving high performance with significantly fewer parameters than conventional transformers.

Researchers significantly boosted the efficiency and quality of rectified flow, a method for generating samples from diffusion models, by introducing novel training techniques that surpass state-of-th…

DMD2 dramatically speeds up image generation by cleverly distilling expensive diffusion models, achieving state-of-the-art results without sacrificing quality.

Immiscible Diffusion boosts diffusion model training efficiency up to 3x by cleverly assigning noise to images, preventing the mixing of data in noise space and thus improving optimization.

IMAGPose: A unified framework generating high-fidelity person images from single or multiple source images & poses, addressing existing methods’ limitations.

Leveraging image understanding models for image tokenizer training dramatically boosts image generation quality, achieving state-of-the-art results.

aSeqDIP: A new unsupervised image reconstruction method using sequential deep image priors, achieving competitive performance with fewer data needs and faster runtimes.

ICDiff, a novel Image Copy Detection system, tackles the unique challenge of identifying replicated content in diffusion model outputs, introducing a specialized dataset and deep embedding method for …

MuDI: a novel framework for multi-subject image personalization, effectively decoupling identities to prevent mixing using segmented subjects and a new evaluation metric.

Researchers solve the conditional image leakage problem in image-to-video diffusion models by proposing a new inference strategy and a time-dependent noise distribution for training. This yields video…

Hyper-SD boosts diffusion model speed by using trajectory segmented consistency distillation and human feedback, achieving state-of-the-art performance.