analiza aiqbench
This arXiv preprint proposes resizing convolutional networks without retraining, by interpolating kernels with standard signal methods (linear, bilinear, bicubic) plus a scaling factor. The authors report usable beyond-training-resolution generation with Stable Diffusion 1.5 and a worst-case ~7% ImageNet accuracy drop, plus ~9% parameter reduction for VGG-16 at ~5% accuracy cost. These are self-reported abstract results that have not been peer-reviewed, so the practical value depends on how the method holds up in independent replication.
Supersampling Stable Diffusion and Beyond: Training-Free Neural Network Scaling via Signal Interpolation
arXiv:2605.08698v3 Announce Type: replace-cross Abstract: This work presents a training-free approach for interpolating neural networks using signal interpolation methods (i.e., linear, bilinear, bicubic). We discovered this phenomenon while attempting to interpolate the convolution layers of Stable-Diffusion-1.5 for beyond-training-resolution (BTR) image generation. As we know, for generating BTR images using DDPMs (Denoising Diffusion Probabilistic Models), there are mainly two approaches, namely, 1) patch-wise diffusion and 2) convolution-kernel dilation. Patch-wise diffusion cannot solve the object-duplication problem efficiently, while the kernel dilation approach may be harder to fine-tune later. Thus, we proposed convolution-kernel interpolation of the Stable Diffusion 1.5 model for $4\times$ BTR image generation. We discover and show mathematically that signal interpolation methods can correctly scale convolution kernels when multiplied by a flat-attenuation factor, while achieving competitive performance against the state of the art. We then applied the same methodology to non-DDPM deep neural network architectures, adapting them for BTR image classification, achieving a worst-case performance drop of $\sim 7\%$ on the ImageNet-1k validation set. Again, we subsample the penultimate fully connected layer of VGG-16 using our method, reducing up to $\sim 9\%$ parameters with $\sim 5\%$ loss in performance. Finally, we also discuss how we can reduce the memory footprint of training neural networks to $4\times$ or greater.
Abstract:This work presents a training-free approach for interpolating neural networks using signal interpolation methods (i.e., linear, bilinear, bicubic). We discovered this phenomenon while attempting to interpolate the convolution layers of Stable-Diffusion-1.5 for beyond-training-resolution (BTR) image generation. As we know, for generating BTR images using DDPMs (Denoising Diffusion Probabilistic Models), there are mainly two approaches, namely, 1) patch-wise diffusion and 2) convolution-kernel dilation. Patch-wise diffusion cannot solve the object-duplication problem efficiently, while the kernel dilation approach may be harder to fine-tune later. Thus, we proposed convolution-kernel interpolation of the Stable Diffusion 1.5 model for $4\times$ BTR image generation. We discover and show mathematically that signal interpolation methods can correctly scale convolution kernels when multiplied by a flat-attenuation factor, while achieving competitive performance against the state of the art. We then applied the same methodology to non-DDPM deep neural network architectures, adapting them for BTR image classification, achieving a worst-case performance drop of $\sim 7\%$ on the ImageNet-1k validation set. Again, we subsample the penultimate fully connected layer of VGG-16 using our method, reducing up to $\sim 9\%$ parameters with $\sim 5\%$ loss in performance. Finally, we also discuss how we can reduce the memory footprint of training neural networks to $4\times$ or greater.
Updated the title for clarity. Removed background and redundant text from section 4.2,5. Improved organization in section 4 and clarity of text in Section 4.3
Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG)
arXiv:2605.08698 [cs.CV]
(or arXiv:2605.08698v3 [cs.CV] for this version)
From: Md. Abu Obaida Zishan [view email] [v1] Sat, 9 May 2026 05:13:21 UTC (24,940 KB) [v2] Thu, 14 May 2026 14:25:41 UTC (24,937 KB) [v3] Tue, 3 Nov 2026 10:19:54 UTC (22,546 KB)
Text integral sub licența deschisă a publicației · 278 cuvinte