Face photo-sketch translation method based on global-local fast Fourier convolution and adaptive cross-domain attention

Multi-condition guided diffusion model for face sketch-to-photo synthesis.

A diffusion-based framework with a stage-wise multi-condition guidance mechanism that enhances both structural and textural fidelity and compares with recent image-to-image translation and diffusion-based baselines to observe competitive performance in both visual coherence and identity preservation.

A Text-Guided Cross-Modal Diffusion Framework With Attention-Based Hand-Drawn Sketches for Face Synthesis

Face sketch-to-photo synthesis plays a key role in computer vision with applications in law enforcement, digital entertainment, and human–computer interaction. Existing generative adversarial network-based methods typically face mode collapse, training instability, and poor performance across different sketching styles. This study introduces Sketch-to-Face, a cross-modal diffusion-based model that leverages Stable Diffusion to generate photorealistic faces using sketch-based inpainting. The proposed approach comprises three components: a Sketch Encoder with multiresolution attention that produces CLIP-compatible embeddings from grayscale sketches, Cross-Modal Fusion module employing bidirectional attention to bridge sketch spatial features with text semantic features, and Automatic Mask Generator with learnable refinement for adaptive inpainting guidance. A low-rank adaptation was applied to fine-tune the U-Net, reducing trainable parameters to 20.5 M while retaining the 860M-parameter backbone. Experiments on the Person Face Sketches dataset (21K+ pairs) show stable convergence and are evaluated using SSIM, LPIPS, FID, identity similarity, runtime, and GPU memory. The results demonstrate a favorable quality–efficiency trade-off compared with text-and-sketch diffusion baselines, while qualitative results indicate plausible facial structure and visual realism.

A Transformer-Based Deformable Convolution UNet for Adaptive Arbitrary Style Transfer

In recent years, image style transfer has matured significantly in the field of computer vision. However, current methods for image style transfer still face the challenge of balancing between content edge contours and style texture strokes, as it is difficult to control the degree of stylization. To address this issue, we propose a UNet based on Transformer feature fusion, named Trans-DCUNet. The network adaptively integrates content and style features by taking advantage of the self-learning characteristics of the cross-attention mechanism in the Transformer. The network combines Transformer and CNN, which takes advantage of the global context capture ability of Transformers and the local modeling ability of convolutional neural networks to fully learn image features at multiple levels, thereby generating stylized images. To enhance texture generation, we introduce a deformable convolutional residual module, which allows the convolution kernel to adapt to varying image features, capturing fine texture details more effectively. Additionally, we augment the traditional perception loss with edge detection loss and frequency perception loss, aiming to better preserve the edge contours of the content image and learn the texture strokes of the style image. Our experiments were conducted on the Microsoft COCO and WikiArt datasets. Experimental results show that our method achieves a content retention SSIM of up to 0.8655 and a style similarity LPIPS of 0.5655, outperforming most competing methods, while generating more artistic stylized images with significantly improved visual effects.

AdaMultiGAN: an adaptive multiscale decoding framework for few-shot image generation

An adaptive multi-scale decoding framework that effectively balances global context with fine-grained detail is proposed that exhibits superior robustness and generalization across diverse domains, effectively alleviating limitations of existing fusion-based approaches.

A Three-Dimensional Novel View Synthesis Method Based on Vision Transformer and Dual Residual Branches

To address the problems of insufficient global structural consistency and local texture blurring in novel view synthesis under single-view conditions, this paper proposes a three-dimensional novel view synthesis method based on the fusion of a Vision Transformer and dual residual branches. The proposed method employs a Vision Transformer (ViT) to extract global features and capture long-range dependencies through a self-attention mechanism. Meanwhile, two complementary local branches are constructed. The RESFB module is designed based on Fast Fourier Convolution to fuse spatial-domain and frequency-domain information, while the RESTiedSE module introduces a TiedSE attention mechanism into the Res2Net framework to adaptively enhance key channel responses. The global and local features are fused in a multi-scale manner and combined with the NeRF volume rendering paradigm to generate novel views. Experiments conducted on the SRN-Chair and SRN-Car subsets of the ShapeNet dataset demonstrate the effectiveness of the proposed method. The results show that the proposed model achieves state-of-the-art SSIM and competitive PSNR, and effectively improves the clarity and structural fidelity of single-view novel view synthesis.

FPSR-GAN: fourier-guided progressive spatial refinement for high-fidelity thangka inpainting