Training-free Regional Prompting for Diffusion Transformers

Diffusion models have demonstrated excellent capabilities in text-to-image generation. Their semantic understanding (i.e., prompt following) ability has also been greatly improved with large language models (e.g., T5, Llama). However, existing models cannot perfectly handle long and complex text prompts, especially when the text prompts contain various objects with numerous attributes and interrelated spatial relationships. While many regional prompting methods have been proposed for UNet-based models (SD1.5, SDXL), but there are still no implementations based on the recent Diffusion Transformer (DiT) architecture, such as SD3 and FLUX.1.In this report, we propose and implement regional prompting for FLUX.1 based on attention manipulation, which enables DiT with fined-grained compositional text-to-image generation capability in a training-free manner. Code is available at https://github.com/antonioo-c/Regional-Prompting-FLUX.

Generalizable Human Gaussians for Sparse View Synthesis

Recent progress in neural rendering has brought forth pioneering methods, such as NeRF and Gaussian Splatting, which revolutionize view rendering across various domains like AR/VR, gaming, and content creation. While these methods excel at interpolating {\em within the training data}, the challenge of generalizing to new scenes and objects from very sparse views persists. Specifically, modeling 3D humans from sparse views presents formidable hurdles due to the inherent complexity of human geometry, resulting in inaccurate reconstructions of geometry and textures. To tackle this challenge, this paper leverages recent advancements in Gaussian Splatting and introduces a new method to learn generalizable human Gaussians that allows photorealistic and accurate view-rendering of a new human subject from a limited set of sparse views in a feed-forward manner. A pivotal innovation of our approach involves reformulating the learning of 3D Gaussian parameters into a regression process defined on the 2D UV space of a human template, which allows leveraging the strong geometry prior and the advantages of 2D convolutions. In addition, a multi-scaffold is proposed to effectively represent the offset details. Our method outperforms recent methods on both within-dataset generalization as well as cross-dataset generalization settings.

Personalized Face Inpainting with Diffusion Models by Parallel Visual Attention

Face inpainting is important in various applications, such as photo restoration, image editing, and virtual reality. Despite the significant advances in face generative models, ensuring that a person's unique facial identity is maintained during the inpainting process is still an elusive goal. Current state-of-the-art techniques, exemplified by MyStyle, necessitate resource-intensive fine-tuning and a substantial number of images for each new identity. Furthermore, existing methods often fall short in accommodating user-specified semantic attributes, such as beard or expression. To improve inpainting results, and reduce the computational complexity during inference, this paper proposes the use of Parallel Visual Attention (PVA) in conjunction with diffusion models. Specifically, we insert parallel attention matrices to each cross-attention module in the denoising network, which attends to features extracted from reference images by an identity encoder. We train the added attention modules and identity encoder on CelebAHQ-IDI, a dataset proposed for identity-preserving face inpainting. Experiments demonstrate that PVA attains unparalleled identity resemblance in both face inpainting and face inpainting with language guidance tasks, in comparison to various benchmarks, including MyStyle, Paint by Example, and Custom Diffusion. Our findings reveal that PVA ensures good identity preservation while offering effective language-controllability. Additionally, in contrast to Custom Diffusion, PVA requires just 40 fine-tuning steps for each new identity, which translates to a significant speed increase of over 20 times.

PATMAT: Person Aware Tuning of Mask-Aware Transformer for Face Inpainting

Generative models such as StyleGAN2 and Stable Diffusion have achieved state-of-the-art performance in computer vision tasks such as image synthesis, inpainting, and de-noising. However, current generative models for face inpainting often fail to preserve fine facial details and the identity of the person, despite creating aesthetically convincing image structures and textures. In this work, we propose Person Aware Tuning (PAT) of Mask-Aware Transformer (MAT) for face inpainting, which addresses this issue. Our proposed method, PATMAT, effectively preserves identity by incorporating reference images of a subject and fine-tuning a MAT architecture trained on faces. By using ~40 reference images, PATMAT creates anchor points in MAT's style module, and tunes the model using the fixed anchors to adapt the model to a new face identity. Moreover, PATMAT's use of multiple images per anchor during training allows the model to use fewer reference images than competing methods. We demonstrate that PATMAT outperforms state-of-the-art models in terms of image quality, the preservation of person-specific details, and the identity of the subject. Our results suggest that PATMAT can be a promising approach for improving the quality of personalized face inpainting.

Extracting Semantic Knowledge from GANs with Unsupervised Learning

Recently, unsupervised learning has made impressive progress on various tasks. Despite the dominance of discriminative models, increasing attention is drawn to representations learned by generative models and in particular, Generative Adversarial Networks (GANs). Previous works on the interpretation of GANs reveal that GANs encode semantics in feature maps in a linearly separable form. In this work, we further find that GAN's features can be well clustered with the linear separability assumption. We propose a novel clustering algorithm, named KLiSH, which leverages the linear separability to cluster GAN's features. KLiSH succeeds in extracting fine-grained semantics of GANs trained on datasets of various objects, e.g., car, portrait, animals, and so on. With KLiSH, we can sample images from GANs along with their segmentation masks and synthesize paired image-segmentation datasets. Using the synthesized datasets, we enable two downstream applications. First, we train semantic segmentation networks on these datasets and test them on real images, realizing unsupervised semantic segmentation. Second, we train image-to-image translation networks on the synthesized datasets, enabling semantic-conditional image synthesis without human annotations.

Bridging the Gap Between Training and Inference of Bayesian Controllable Language Models

Large-scale pre-trained language models have achieved great success on natural language generation tasks. However, it is difficult to control the pre-trained language models to generate sentences with the desired attribute such as topic and sentiment, etc. Recently, Bayesian Controllable Language Models (BCLMs) have been shown to be efficient in controllable language generation. Rather than fine-tuning the parameters of pre-trained language models, BCLMs use external discriminators to guide the generation of pre-trained language models. However, the mismatch between training and inference of BCLMs limits the performance of the models. To address the problem, in this work we propose a "Gemini Discriminator" for controllable language generation which alleviates the mismatch problem with a small computational cost. We tested our method on two controllable language generation tasks: sentiment control and topic control. On both tasks, our method reached achieved new state-of-the-art results in automatic and human evaluations.

Linear Semantics in Generative Adversarial Networks

Generative Adversarial Networks (GANs) are able to generate high-quality images, but it remains difficult to explicitly specify the semantics of synthesized images. In this work, we aim to better understand the semantic representation of GANs, and thereby enable semantic control in GAN's generation process. Interestingly, we find that a well-trained GAN encodes image semantics in its internal feature maps in a surprisingly simple way: a linear transformation of feature maps suffices to extract the generated image semantics. To verify this simplicity, we conduct extensive experiments on various GANs and datasets; and thanks to this simplicity, we are able to learn a semantic segmentation model for a trained GAN from a small number (e.g., 8) of labeled images. Last but not least, leveraging our findings, we propose two few-shot image editing approaches, namely Semantic-Conditional Sampling and Semantic Image Editing. Given a trained GAN and as few as eight semantic annotations, the user is able to generate diverse images subject to a user-provided semantic layout, and control the synthesized image semantics. We have made the code publicly available.

Frame Difference-Based Temporal Loss for Video Stylization

Neural style transfer models have been used to stylize an ordinary video to specific styles. To ensure temporal inconsistency between the frames of the stylized video, a common approach is to estimate the optic flow of the pixels in the original video and make the generated pixels match the estimated optical flow. This is achieved by minimizing an optical flow-based (OFB) loss during model training. However, optical flow estimation is itself a challenging task, particularly in complex scenes. In addition, it incurs a high computational cost. We propose a much simpler temporal loss called the frame difference-based (FDB) loss to solve the temporal inconsistency problem. It is defined as the distance between the difference between the stylized frames and the difference between the original frames. The differences between the two frames are measured in both the pixel space and the feature space specified by the convolutional neural networks. A set of human behavior experiments involving 62 subjects with 25,600 votes showed that the performance of the proposed FDB loss matched that of the OFB loss. The performance was measured by subjective evaluation of stability and stylization quality of the generated videos on two typical video stylization models. The results suggest that the proposed FDB loss is a strong alternative to the commonly used OFB loss for video stylization.