In this paper, we propose a novel implicit neural representation (INR)-based image compression paradigm that generate image-specific latents from a multi-scale Gaussian noise tensor which is deterministically generated using a shared random seed. As shown in Fig. 1, inspired by the normalization strategy in E2E codecs where latents are transformed into Gaussian noise to demonstrate the removal of spatial redundancy, we explore the inverse direction: generating Gaussian latents directly from Gaussian noise. Specifically, given a noise input$z$, we introduce a Gaussian Parameter Prediction (GPP) module that estimates the Gaussian distribution parameters, enabling one-shot latent generation via the reparameterization trick. The predicted latent is then transformed into the reconstructed image through a synthesis network.

Recently, implicit neural representation (INR)-based image compression has attracted growing attention due to its low decoding complexity. However, most existing methods require multiple encoding processes to generate multiple bitstreams and reconstructions at different quality levels, leading to high storage and computational overhead. A more practical solution is progressive compression, which supports multiple qualities within a single bitstream. In this paper, we propose PINIC (Progressive Implicit Neural Image Compression), a DWT-guided framework that integrates progressive functionality into INR-based compression for the first time. Specifically, DWT energy is applied to adaptively allocate the energy of latent channels and latent will be entropy coded in the order of channel energy. Moreover, we adopt reparameterization to decouple network structures between training and inference, retaining the expressive power of complex architectures while reducing parameter count at deployment. Extensive experiments demonstrate that PINIC achieves state-of-the-art rate-distortion performance on the Kodak and CLIC2020P datasets, while simultaneously supporting progressive compression.

This study addresses the challenge of domain adaptation in learned image compression, focusing on shifting the model from natural images to user-generated content (UGC) domain. We propose a novel entropy adapter framework augmented with knowledge distillation techniques to improve performance. Unlike existing adapter-based methods that primarily enhance transformation modules, we identify the mismatch between the adapter-based transformation and the fixed entropy network. To resolve this, we introduce adapters within the hypernet and entropy model. Specifically, our decoupled entropy adapter features a deeper residual structure with two independent branches, enabling a separate refinement of mean and scale components. This design improves the accuracy of probability estimation and overall compression efficiency. To further enhance the effectiveness of the adapters, we incorporate a knowledge distillation (KD) strategy with a progressive loss function. It facilitates a smooth transition from KD loss to a rate-distortion (RD) loss in the training process, effectively transferring knowledge from a directly fine-tuned model to the student model. Consequently, this strengthens the adapter's learning capability and improves compression performance. Experimental results show that the proposed method achieves a significant 11.5% bitrate savings compared to the baseline model. Additionally, it demonstrates robust adaptability across diverse network architectures.

In this paper, we present a Conv-Attention, a decoder-friendly attention mechanism, in an effort to advancing the practical application of the artificial intelligence-based image coding. More specifically, the proposed method is tailored for JPEG AI, which is the latest advanced neural-network based image coding standard. By identifying the obstacles by profiling the decoding complexity of JPEG AI, the attention module accounts for a significant proportion, which mainly attributes to the intricate network structure and involvement of less efficient operations. Conv-Attention model is composed with plain convolution and activation computations, equipping with sub-scaling and up-scaling design, such that the non-adjacent features can be well captured, leading to the reduction of decoding complexity and maintenance of the synthesis and attentive capability. Simulation results verify the effectiveness of the proposed method with JPEG AI reference software, wherein the decoding complexity is reduced by 80% with negligible coding performance loss. The proposed method was adopted in the 100th JPEG meeting.