Hang Guo (郭航)

Autoregressive video diffusion models have shown impressive generation quality but suffer from high computational cost, particularly in the attention layers over long frame sequences. We find that approximately 25% of attention heads act as “dummy heads” that attend almost exclusively to the current frame, contributing little to temporal modeling. Based on this insight, we propose Dummy Forcing, a training-free method that identifies and exploits these dummy heads for efficient autoregressive video generation. Specifically, we introduce heterogeneous memory allocation that reduces head-wise context redundancy, dynamic head programming that adaptively classifies head types, and context packing that enables aggressive cache compression. Experiments demonstrate up to 2.0× speedup with <0.5% quality degradation, achieving 24.3 FPS real-time video generation at 720P resolution.

Large Language Models (LLMs) have achieved remarkable success but their massive size poses significant challenges for practical deployment. While quantization and sparsification are two mainstream compression techniques, jointly applying them remains challenging due to their conflicting requirements: quantization prefers compact weight distributions while pruning benefits from high variance. In this work, we propose Optimal Brain Restoration (OBR), a training-free framework for joint quantization and sparsification of LLMs. OBR formulates the compression as a second-order Hessian-based optimization problem, which we reformulate into a tractable surrogate approximation with a closed-form solution via group error compensation. This enables simultaneous W4A4KV4 quantization and 50% sparsity on LLMs, achieving up to 4.72× speedup and 6.4× memory reduction compared to FP16-dense baselines. Extensive experiments on Llama-2, Llama-3, and Qwen2.5 families validate the effectiveness of OBR.

Visual Autoregressive (VAR) modeling has gained popularity for its shift towards next-scale prediction. However, existing VAR paradigms process the entire token map at each scale step, leading to the complexity and runtime scaling dramatically with image resolution. To address this challenge, we propose FastVAR, a post-training acceleration method for efficient resolution scaling with VARs. Our key finding is that the majority of latency arises from the large-scale step where most tokens have already converged. Leveraging this observation, we develop the cached token pruning strategy that only forwards pivotal tokens for scale-specific modeling while using cached tokens from previous scale steps to restore the pruned slots. This significantly reduces the number of forwarded tokens and improves the efficiency at larger resolutions. Experiments show the proposed FastVAR can further speedup FlashAttention-accelerated VAR by 2.7 with negligible performance drop of <1%. We further extend FastVAR to zero-shot generation of higher resolution images. In particular, FastVAR can generate one 2K image with 15GB memory footprints in 1.5s on a single NVIDIA 3090 GPU. Code is available at https://github.com/csguoh/FastVAR.

The Mamba-based image restoration backbones have recently demonstrated significant potential in balancing global reception and computational efficiency. However, the inherent causal modeling limitation of Mamba, where each token depends solely on its predecessors in the scanned sequence, restricts the full utilization of pixels across the image and thus presents new challenges in image restoration. In this work, we propose MambaIRv2, which equips Mamba with the non-causal modeling ability similar to ViTs to reach the attentive state space restoration model. Specifically, the proposed attentive state-space equation allows to attend beyond the scanned sequence and facilitate image unfolding with just one single scan. Moreover, we further introduce a semantic-guided neighboring mechanism to encourage interaction between distant but similar pixels. Extensive experiments show our MambaIRv2 outperforms SRFormer by even 0.35dB PSNR for lightweight SR even with 9.3% less parameters and suppresses HAT on classic SR by up to 0.29dB.

Fine-tuning large-scale text-to-image diffusion models for various downstream tasks has yielded impressive results. However, the heavy computational burdens of tuning large models prevent personal customization. Recent advances have attempted to employ parameter-efficient fine-tuning (PEFT) techniques to adapt the floating-point (FP) or quantized pre-trained weights. Nonetheless, the adaptation parameters in existing works are still restricted to FP arithmetic, hindering hardware-friendly acceleration. In this work, we propose IntLoRA, to further push the efficiency limits by using integer type (INT) low-rank parameters to adapt the quantized diffusion models. By working in the integer arithmetic, our IntLoRA offers three key advantages: (i) for fine-tuning, the pre-trained weights are quantized, reducing memory usage; (ii) for storage, both pre-trained and low-rank weights are in INT which consumes less disk space; (iii) for inference, IntLoRA weights can be naturally merged into quantized pre-trained weights through efficient integer multiplication or bit-shifting, eliminating additional post-training quantization. Extensive experiments demonstrate that IntLoRA can achieve performance on par with or even superior to the vanilla LoRA, accompanied by significant efficiency improvements.

Pre-training has shown promising results on various image restoration tasks, which is usually followed by full fine-tuning for each specific downstream task (e.g., image denoising). However, such full fine-tuning usually suffers from the problems of heavy computational cost in practice, due to the massive parameters of pre-trained restoration models, thus limiting its real-world applications. Recently, Parameter Efficient Transfer Learning (PETL) offers an efficient alternative solution to full fine-tuning, yet still faces great challenges for pre-trained image restoration models, due to the diversity of different degradations. To address these issues, we propose AdaptIR, a novel parameter efficient transfer learning method for adapting pre-trained restoration models. Specifically, the proposed method consists of a multi-branch inception structure to orthogonally capture local spatial, global spatial, and channel interactions. In this way, it allows powerful representations under a very low parameter budget. Extensive experiments demonstrate that the proposed method can achieve comparable or even better performance than full fine-tuning, while only using 0.6% parameters.

Recent advances in diffusion-based Large Restoration Models (LRMs) have significantly improved photo-realistic image restoration by leveraging the internal knowledge embedded within model weights. However, existing LRMs often suffer from the hallucination dilemma, i.e., producing incorrect contents or textures when dealing with severe degradations, due to their heavy reliance on limited internal knowledge. In this paper, we propose an orthogonal solution called the Retrieval-augmented Framework for Image Restoration (ReFIR), which incorporates retrieved images as external knowledge to extend the knowledge boundary of existing LRMs in generating details faithful to the original scene. Specifically, we first introduce the nearest neighbor lookup to retrieve content-relevant high-quality images as reference, after which we propose the cross-image injection to modify existing LRMs to utilize high-quality textures from retrieved images. Thanks to the additional external knowledge, our ReFIR can well handle the hallucination challenge and facilitate faithfully results. Extensive experiments demonstrate that ReFIR can achieve not only high-fidelity but also realistic restoration results. Importantly, our ReFIR requires no training and is adaptable to various LRMs.

Recent years have witnessed great progress in image restoration thanks to the advancements in modern deep neural networks e.g. Convolutional Neural Network and Transformer. However, existing restoration backbones are usually limited due to the inherent local reductive bias or quadratic computational complexity. Recently, Selective Structured State Space Model e.g., Mamba, have shown great potential for long-range dependencies modeling with linear complexity, but it is still under-explored in low-level computer vision. In this work, we introduce a simple but strong benchmark model, named MambaIR, for image restoration. In detail, we propose the Residual State Space Block as the core component, which employs convolution and channel attention to enhance capabilities of the vanilla Mamba. In this way, our MambaIR takes advantages of local patch recurrence prior as well as channel interaction to produce restoration-specific feature representation. Extensive experiments demonstrate the superiority of our method, for example, MambaIR outperforms Transformer-based baseline SwinIR by up to 0.36dB, using similar computational cost but with global receptive field.