Pre-trained visual models enjoy strong representations, yet suffer from massive parameters to be shifted in downstream practices. Many parameter-efficient fine-tuning methods have been proposed, mostly requiring only 1% additional parameters to achieve comparable results. However, current solutions either consider all feature channels equally or detect saliencies with individual layer, leading to many redundancies reserved. To address current issues, this paper proposes a new parameter fine-tuning method named “Gradient Selection Tuning” (GST), which leverages gradients that are capable of capturing the cascading effects across successive channels. Instead of saliency detection, we turn to compress the redundancies for channel selection, since the computed gradient values enjoy much lower mutual information. With GST facilitated, we further elaborate an Information-Guided Adapter following information bottleneck theory, effectively performing parameter compression yet with task-specific features preserved. Experimental results demonstrate that our method outperforms the baseline methods by adding only 0.075M parameters to ViT-B backbone. On domain generalization, our proposal also enjoys strong performance in low-parameter scenarios.

Recent advances in pretrained diffusion models, particularly the FreeControl, have enabled fine-grained spatial control in text-to-image generation. However, FreeControl still suffers notable limitations in detail generation and appearance synthesis. With deep analysis performed, we reveal that inadequate feature representation in the early generation phases is the main cause of the insufficient structure elaboration. To address this issue, we dig into the temporal evolution of inverse attention features, then extract more expressive structure information as the input to the guidance function, ensuring formation integrity. Moreover, to tackle the surface degradation and ambiguity caused by the dual guidance of structure and appearance, we engineer the Adaptive Instance Normalization (AdaIN) mechanism into a latent space, rather than the typical feature space, during the intermediate generation stage. This improvement not only guarantees the close alignment between the generated image and structural reference but also significantly strengthens the appearance modeling capability and optimizes the texture representation of both foreground and background elements. Extensive experiments demonstrate that our proposal consistently outperforms existing baseline models across multiple metrics, including Self-sim, CLIP, and LPIPS. Both quantitative and qualitative results confirm that our approach achieves superior performance in terms of content consistency, visual quality, and detail preservation.

Benefiting from the booming deep learning techniques, neural operators (NO) are considered as an ideal alternative to break the traditions of solving Partial Differential Equations (PDE) with expensive cost. Yet with the remarkable progress, current solutions concern little on the holistic function features--both global and local information-- during the process of solving PDEs. Besides, a meticulously designed kernel integration to meet desirable performance often suffers from a severe computational burden, such as GNO with $O(N(N-1))$, FNO with $O(NlogN)$, and Transformer-based NO with $O(N^2)$. To counteract the dilemma, we propose a mamba neural operator with $O(N)$ computational complexity, namely MambaNO. Functionally, MambaNO achieves a clever balance between global integration, facilitated by state space model of Mamba that scans the entire function, and local integration, engaged with an alias-free architecture. We prove a property of continuous-discrete equivalence to show the capability of MambaNO in approximating operators arising from universal PDEs to desired accuracy. MambaNOs are evaluated on a diverse set of benchmarks with possibly multi-scale solutions and set new state-of-the-art scores, yet with fewer parameters and better efficiency.