COAP: Memory-Efficient Training with Correlation-Aware Gradient Projection

Training large-scale neural networks in vision, and multimodal domains demands substantial memory resources, primarily due to the storage of optimizer states. While LoRA, a popular parameter-efficient method, reduces memory usage, it often suffers from suboptimal performance due to the constraints of low-rank updates. Low-rank gradient projection methods (e.g., GaLore, Flora) reduce optimizer memory by projecting gradients and moment estimates into low-rank spaces via singular value decomposition or random projection. However, they fail to account for inter-projection correlation, causing performance degradation, and their projection strategies often incur high computational costs. In this paper, we present COAP (Correlation-Aware Gradient Projection), a memory-efficient method that minimizes computational overhead while maintaining training performance. Evaluated across various vision, language, and multimodal tasks, COAP outperforms existing methods in both training speed and model performance. For LLaMA-1B, it reduces optimizer memory by 61% with only 2% additional time cost, achieving the same PPL as AdamW. With 8-bit quantization, COAP cuts optimizer memory by 81% and achieves 4x speedup over GaLore for LLaVA-v1.5-7B fine-tuning, while delivering higher accuracy.

ID-Patch: Robust ID Association for Group Photo Personalization

The ability to synthesize personalized group photos and specify the positions of each identity offers immense creative potential. While such imagery can be visually appealing, it presents significant challenges for existing technologies. A persistent issue is identity (ID) leakage, where injected facial features interfere with one another, resulting in low face resemblance, incorrect positioning, and visual artifacts. Existing methods suffer from limitations such as the reliance on segmentation models, increased runtime, or a high probability of ID leakage. To address these challenges, we propose ID-Patch, a novel method that provides robust association between identities and 2D positions. Our approach generates an ID patch and ID embeddings from the same facial features: the ID patch is positioned on the conditional image for precise spatial control, while the ID embeddings integrate with text embeddings to ensure high resemblance. Experimental results demonstrate that ID-Patch surpasses baseline methods across metrics, such as face ID resemblance, ID-position association accuracy, and generation efficiency. Project Page is: https://byteaigc.github.io/ID-Patch/

MoMA: Multimodal LLM Adapter for Fast Personalized Image Generation

In this paper, we present MoMA: an open-vocabulary, training-free personalized image model that boasts flexible zero-shot capabilities. As foundational text-to-image models rapidly evolve, the demand for robust image-to-image translation grows. Addressing this need, MoMA specializes in subject-driven personalized image generation. Utilizing an open-source, Multimodal Large Language Model (MLLM), we train MoMA to serve a dual role as both a feature extractor and a generator. This approach effectively synergizes reference image and text prompt information to produce valuable image features, facilitating an image diffusion model. To better leverage the generated features, we further introduce a novel self-attention shortcut method that efficiently transfers image features to an image diffusion model, improving the resemblance of the target object in generated images. Remarkably, as a tuning-free plug-and-play module, our model requires only a single reference image and outperforms existing methods in generating images with high detail fidelity, enhanced identity-preservation and prompt faithfulness. Our work is open-source, thereby providing universal access to these advancements.

MagicDance: Realistic Human Dance Video Generation with Motions & Facial Expressions Transfer

In this work, we propose MagicDance, a diffusion-based model for 2D human motion and facial expression transfer on challenging human dance videos. Specifically, we aim to generate human dance videos of any target identity driven by novel pose sequences while keeping the identity unchanged. To this end, we propose a two-stage training strategy to disentangle human motions and appearance (e.g., facial expressions, skin tone and dressing), consisting of the pretraining of an appearance-control block and fine-tuning of an appearance-pose-joint-control block over human dance poses of the same dataset. Our novel design enables robust appearance control with temporally consistent upper body, facial attributes, and even background. The model also generalizes well on unseen human identities and complex motion sequences without the need for any fine-tuning with additional data with diverse human attributes by leveraging the prior knowledge of image diffusion models. Moreover, the proposed model is easy to use and can be considered as a plug-in module/extension to Stable Diffusion. We also demonstrate the model's ability for zero-shot 2D animation generation, enabling not only the appearance transfer from one identity to another but also allowing for cartoon-like stylization given only pose inputs. Extensive experiments demonstrate our superior performance on the TikTok dataset.

Do We Really Need to Learn Representations from In-domain Data for Outlier Detection?

Unsupervised outlier detection, which predicts if a test sample is an outlier or not using only the information from unlabelled inlier data, is an important but challenging task. Recently, methods based on the two-stage framework achieve state-of-the-art performance on this task. The framework leverages self-supervised representation learning algorithms to train a feature extractor on inlier data, and applies a simple outlier detector in the feature space. In this paper, we explore the possibility of avoiding the high cost of training a distinct representation for each outlier detection task, and instead using a single pre-trained network as the universal feature extractor regardless of the source of in-domain data. In particular, we replace the task-specific feature extractor by one network pre-trained on ImageNet with a self-supervised loss. In experiments, we demonstrate competitive or better performance on a variety of outlier detection benchmarks compared with previous two-stage methods, suggesting that learning representations from in-domain data may be unnecessary for outlier detection.

EBMs Trained with Maximum Likelihood are Generator Models Trained with a Self-adverserial Loss

Maximum likelihood estimation is widely used in training Energy-based models (EBMs). Training requires samples from an unnormalized distribution, which is usually intractable, and in practice, these are obtained by MCMC algorithms such as Langevin dynamics. However, since MCMC in high-dimensional space converges extremely slowly, the current understanding of maximum likelihood training, which assumes approximate samples from the model can be drawn, is problematic. In this paper, we try to understand this training procedure by replacing Langevin dynamics with deterministic solutions of the associated gradient descent ODE. Doing so allows us to study the density induced by the dynamics (if the dynamics are invertible), and connect with GANs by treating the dynamics as generator models, the initial values as latent variables and the loss as optimizing a critic defined by the very same energy that determines the generator through its gradient. Hence the term - self-adversarial loss. We show that reintroducing the noise in the dynamics does not lead to a qualitative change in the behavior, and merely reduces the quality of the generator. We thus show that EBM training is effectively a self-adversarial procedure rather than maximum likelihood estimation.

Exponential Tilting of Generative Models: Improving Sample Quality by Training and Sampling from Latent Energy

In this paper, we present a general method that can improve the sample quality of pre-trained likelihood based generative models. Our method constructs an energy function on the latent variable space that yields an energy function on samples produced by the pre-trained generative model. The energy based model is efficiently trained by maximizing the data likelihood, and after training, new samples in the latent space are generated from the energy based model and passed through the generator to producing samples in observation space. We show that using our proposed method, we can greatly improve the sample quality of popular likelihood based generative models, such as normalizing flows and VAEs, with very little computational overhead.

Likelihood Regret: An Out-of-Distribution Detection Score For Variational Auto-encoder

Deep probabilistic generative models enable modeling the likelihoods of very high dimensional data. An important application of generative modeling should be the ability to detect out-of-distribution (OOD) samples by setting a threshold on the likelihood. However, a recent study shows that probabilistic generative models can, in some cases, assign higher likelihoods on certain types of OOD samples, making the OOD detection rules based on likelihood threshold problematic. To address this issue, several OOD detection methods have been proposed for deep generative models. In this paper, we make the observation that some of these methods fail when applied to generative models based on Variational Auto-encoders (VAE). As an alternative, we propose Likelihood Regret, an efficient OOD score for VAEs. We benchmark our proposed method over existing approaches, and empirical results suggest that our method obtains the best overall OOD detection performances compared with other OOD method applied on VAE.

A Method to Model Conditional Distributions with Normalizing Flows

In this work, we investigate the use of normalizing flows to model conditional distributions. In particular, we use our proposed method to analyze inverse problems with invertible neural networks by maximizing the posterior likelihood. Our method uses only a single loss and is easy to train. This is an improvement on the previous method that solves similar inverse problems with invertible neural networks but which involves a combination of several loss terms with ad-hoc weighting. In addition, our method provides a natural framework to incorporate conditioning in normalizing flows, and therefore, we can train an invertible network to perform conditional generation. We analyze our method and perform a careful comparison with previous approaches. Simple experiments show the effectiveness of our method, and more comprehensive experimental evaluations are undergoing.

Generative Latent Flow: A Framework for Non-adversarial Image Generation

Generative Adversarial Networks (GANs) have been shown to outperform non-adversarial generative models in terms of the image generation quality by a large margin. Recently, researchers have looked into improving non-adversarial alternatives that can close the gap of generation quality while avoiding some common issues of GANs, such as unstable training and mode collapse. Examples in this direction include Two-stage VAE and Generative Latent Nearest Neighbors. However, a major drawback of these models is that they are slow to train, and in particular, they require two training stages. To address this, we propose Generative Latent Flow (GLF), which uses an auto-encoder to learn the mapping to and from the latent space, and an invertible flow to map the distribution in the latent space to simple i.i.d noise. The advantages of our method include a simple conceptual framework, single stage training and fast convergence. Quantitatively, the generation quality of our model significantly outperforms that of VAEs, and is competitive with GANs' benchmark on commonly used datasets.