Real-time instruction-based portrait image editing is crucial in various applications, including filters, augmented reality, and video communications, etc. However, real-time portrait editing presents three significant challenges: identity preservation, fidelity to editing instructions, and fast model inference. Given that these aspects often present a trade-off, concurrently addressing them poses an even greater challenge. While diffusion-based image editing methods have shown promising capabilities in personalized image editing in recent years, they lack a dedicated focus on portrait editing and thus suffer from the aforementioned problems as well. To address the gap, this paper introduces an Instant-Portrait Network (IPNet), the first one-step diffusion-based model for portrait editing. We train the network in two stages. We first employ an annealing identity loss to train an Identity Enhancement Network (IDE-Net), to ensure robust identity preservation. We then train the IPNet using a novel diffusion Multi-Objective Distillation approach that integrates adversarial loss, identity distillation loss, and a novel Facial-Style Enhancing loss. The Diffusion Multi-Objective Distillation approach efficiently reduces inference steps, ensures identity consistency, and enhances the precision of instruction-based editing. Extensive comparison with prior models demonstrates IPNet as a superior model in terms of identity preservation, text fidelity, and inference speed.

Adapter-based training has emerged as a key mechanism for extending the capabilities of powerful foundation image generators, enabling personalized and stylized text-to-image synthesis. These adapters are typically trained to capture a specific target attribute, such as subject identity, using single-image reconstruction objectives. However, because the input image inevitably contains a mixture of visual factors, adapters are prone to entangle the target attribute with incidental ones, such as pose, expression, and lighting. This spurious correlation problem limits generalization and obstructs the model's ability to adhere to the input text prompt. In this work, we uncover a simple yet effective solution: provide the very shortcuts we wish to eliminate during adapter training. In Shortcut-Rerouted Adapter Training, confounding factors are routed through auxiliary modules, such as ControlNet or LoRA, eliminating the incentive for the adapter to internalize them. The auxiliary modules are then removed during inference. When applied to tasks like facial and full-body identity injection, our approach improves generation quality, diversity, and prompt adherence. These results point to a general design principle in the era of large models: when seeking disentangled representations, the most effective path may be to establish shortcuts for what should NOT be learned.

Diffusion-based text-to-image generative models, e.g., Stable Diffusion, have revolutionized the field of content generation, enabling significant advancements in areas like image editing and video synthesis. Despite their formidable capabilities, these models are not without their limitations. It is still challenging to synthesize an image that aligns well with the input text, and multiple runs with carefully crafted prompts are required to achieve satisfactory results. To mitigate these limitations, numerous studies have endeavored to fine-tune the pre-trained diffusion models, i.e., UNet, utilizing various technologies. Yet, amidst these efforts, a pivotal question of text-to-image diffusion model training has remained largely unexplored: Is it possible and feasible to fine-tune the text encoder to improve the performance of text-to-image diffusion models? Our findings reveal that, instead of replacing the CLIP text encoder used in Stable Diffusion with other large language models, we can enhance it through our proposed fine-tuning approach, TextCraftor, leading to substantial improvements in quantitative benchmarks and human assessments. Interestingly, our technique also empowers controllable image generation through the interpolation of different text encoders fine-tuned with various rewards. We also demonstrate that TextCraftor is orthogonal to UNet finetuning, and can be combined to further improve generative quality.

Diffusion-based image generation models have achieved great success in recent years by showing the capability of synthesizing high-quality content. However, these models contain a huge number of parameters, resulting in a significantly large model size. Saving and transferring them is a major bottleneck for various applications, especially those running on resource-constrained devices. In this work, we develop a novel weight quantization method that quantizes the UNet from Stable Diffusion v1.5 to $1.99$ bits, achieving a model with $7.9\times$ smaller size while exhibiting even better generation quality than the original one. Our approach includes several novel techniques, such as assigning optimal bits to each layer, initializing the quantized model for better performance, and improving the training strategy to dramatically reduce quantization error. Furthermore, we extensively evaluate our quantized model across various benchmark datasets and through human evaluation to demonstrate its superior generation quality.

User interaction provides useful information for solving challenging computer vision problems in practice. In this paper, we show that a very limited number of user clicks could greatly boost monocular depth estimation performance and overcome monocular ambiguities. We formulate this task as a deep structured model, in which the structured pixel-wise depth estimation has ordinal constraints introduced by user clicks. We show that the inference of the proposed model could be efficiently solved through a feed-forward network. We demonstrate the effectiveness of the proposed model on NYU Depth V2 and Stanford 2D-3D datasets. On both datasets, we achieve state-of-the-art performance when encoding user interaction into our deep models.