APISR: Anime Production Inspired Real-World Anime Super-Resolution - Supplementary Material - 训练数据与实现细节

Aug 21, 2026

Upload by Fahsai

APISR provides supplementary technical material for anime-focused image super-resolution, detailing dataset construction, network training choices, and degradation modeling. The document introduces an API SR dataset with 3,740 high-quality images, then specifies generator training architecture adjustments, hand-drawn line enhancement using XDoG-derived edge maps with outlier filtering and passive dilation, and a balanced twin perceptual loss using ResNet50 feature comparisons. It further explains prediction-oriented compression settings across image and video codecs and reports additional comparative and ablation visual results.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages9

Reading Time23 min

AccessFree to download

API Dataset Details

Implementation Details

Training Network Details

Hand-drawn line enhancement Details

Balanced Twin Perceptual Loss Details

Degradation Details

Visual Results

APISR: Anime Production Inspired Real-World Anime Super-Resolution Supplementary Material In this supplementary material, Sec. A first presents more statistics and details of our proposed anime image SR training dataset. Then, Sec. B shows details about our implementations in super-resolution (SR) network training. Specifically, Sec. B.1 presents the image SR network we used in our training. Sec. B.2 presents details of post-processing techniques we use on the pseudo-GT preparation for hand-drawn line enhancement. Sec. B.3 presents figures and details of the ResNet50 [10] perceptual loss for our proposed balanced twin perceptual loss. Sec. B.4 provides the hyperparameter setting for our proposed prediction-oriented compression and shuffled resize module in the degradation model. Finally, Sec. C provides more visual results of comparisons among SOTA methods and ablation studies. A. API Dataset Details Our Anime Production-oriented Image (API) SR dataset contains 3,740 high-quality and informative images. This quantity is roughly the same quantity as the previous photorealistic SR training dataset size [26, 29], which includes DIV2K [1], Flickr2K [20], and OutdoorSceneTraining [24] . The aspect ratio and resolution information before scaling are shown in Fig. 1. B. Implementation Details B.1. Training Network Details The generator network we deploy is GRL [14], a SOTA image SR network (CVPR 2023) . GRL leverages interconnected relationships within various layers of image structures through a Transformer-based framework, attaining improvement in multiple tasks of SR and image restoration. The model we chose is its tiny version, which has 0.91M parameters. To better adapt the real-world SR task, we changed its upsampler module from the default pixel shuffle strategy to the nearest neighbor interpolation with the convolution layer approach, which is used for the base model version but not for the tiny version in their proposed methods. We change the upsampler because the nearest neighbor interpolation with the convolution layer is claimed to show fewer artifacts in the upsampling process than the pixel shuffle strategy. The final network parameter is 1.03M, which is the smallest network among all image and videobased SOTA methods that we compare. B.2. Hand-drawn line enhancement Details In the hand-drawn line enhancement, we have proposed outlier filter and passive dilate techniques to obtain a clean XDoG-extracted [27] hand-drawn line edge map. XDoG is widely used in paired dataset preparation in anime colorization [4, 5, 11, 23] . The extracted edge map by XDoG is a binary output, where the white pixel stands for the active edge map region and the black pixel stands for the unrelated region. For the outlier filter, we use breadth-first search in eight directions to recursively detect the surrounding pixels of all white pixels and turn white pixel regions into black pixels if the total quantity of connected white pixels is less than the threshold. We empirically set the threshold as 32 . For the dilation, we passively replace the black pixel with the white pixel if it has more than 3 white pixel neighbors, which is different from independent kernel-based active dilation methods in [7, 9, 13] that directly spread the surrounding neighbors to be white pixels if the central pixel is white. Compared to active dilation methods, our proposed passive dilation is more concentrated on the handdrawn lines region instead of covering unrelated pixel information (see Fig. 3) . Thus, we name our methods as passive dilatation. In the implementation, we will do an unsharp mask for the whole image first to increase overall visualization sharpness and then apply two extra turns of sharpening to the hand-drawn lines specifically based on the pipeline design mentioned above. More implementation details can be found in our released code. B.3. Balanced Twin Perceptual Loss Details As shown in Fig. 2, our proposed middle-laye

FAQs

APISR 的数据集规模与特点是什么？

API SR 数据集包含 3,740 张高质量且信息量充分的动漫相关图像。文中给出了缩放前的宽高比与分辨率统计，并与先前的真实感超分训练数据集规模进行对比说明。

训练网络在 GRL 上做了哪些关键实现调整？

使用 GRL 的 tiny 版本作为生成器，并将默认的 upsampler（像素洗牌）替换为“最近邻插值 + 卷积层”的组合，以减少上采样过程中的伪影。最终参数规模为 1.03M，并强调其与对比方法相比的紧凑性。

手绘线增强部分如何从 XDoG 得到更干净的线条边缘图？

先用 XDoG 提取二值边缘图，然后通过 outlier filter 以连通白像素数量阈值递归剔除离群区域；再采用 passive dilate 根据白邻居数量对黑像素进行被动替换，从而更集中地扩展手绘线条区域。

退化模型中的预测导向压缩设置如何覆盖图像与视频？

预测导向压缩同时使用带预测机制的图像压缩（WebP、AVIF）以及单帧视频压缩。文中给出 JPEG/WebP/AVIF 与 MPEG2/MPEG4/H.264/H.265 的质量参数范围、CRF 或 qscale 控制策略，并说明 preset 的取值概率分布。

Aug 21, 2026

Upload by Fahsai