Diffusion Models |

Alibaba Group - Tmall Campus

Sun, 01 Mar 2026 00:00:00 +0000

Overview

Algorithm Intern Mar. 2026 – Apr. 2026

Portrait-to-Anime Generation System

Developed portrait stylization models based on StyleGAN2 and diffusion architectures for high-quality anime-style image generation.
Constructed end-to-end training workflows including dataset preprocessing, data augmentation, model training, inference, and quantitative evaluation.
Built scalable pipelines supporting rapid experimentation across multiple generative model architectures.

Model Optimization and Training Stability

Improved training stability and convergence performance using PyTorch and TensorFlow through optimizer tuning, learning-rate scheduling, and architecture refinement.
Conducted systematic hyperparameter exploration to improve image fidelity, visual quality, and generation consistency.
Evaluated generation performance using both objective metrics and human perceptual assessments.

Representation Enhancement

Explored attention-based feature enhancement and LoRA-based parameter-efficient adaptation techniques for controllable style transfer.
Improved preservation of facial identity features while maintaining target anime-style characteristics.
Enhanced style consistency and fine-grained visual details across generated outputs.

Experimental Platform Development

Designed reproducible experiment management pipelines covering data processing, training, inference, and result analysis.
Automated model evaluation and visualization workflows to support efficient iteration and comparison of generative model variants.

PERoKF: Physics-Enhanced Super-Resolution of Kolmogorov Flow

Fri, 19 Dec 2025 00:00:00 +0000

Team

YiFan Cai
YiRan Hu
JiaCheng Zhu

Overview

PERoKF studies how to inject physics priors into modern neural networks for single-frame super-resolution of 2D Kolmogorov flow—without requiring temporal supervision. The goal is to recover fine-scale turbulent structures lost by downsampling.

Input (LR): 128×128 vorticity
Target (HR): 512×512 vorticity (4×)

Methods

We evaluate four model families (all operating on bicubic-upsampled LR inputs and predicting HR vorticity):

CNN: lightweight hierarchical convolutional baseline
UNet: multi-scale ResNet-style UNet (optional attention)
FNO: Fourier Neural Operator with truncated spectral modes
Diffusion: conditional DDPM with UNet backbone (x₀-prediction)

Physics-guided strategies

Physics-derived feature augmentation

Laplacian, streamfunction, velocity components, nonlinear advection terms

Physics-consistency loss

Navier–Stokes residual loss using the same pseudo-spectral operator as data generation
Enforces consistency via implied time derivatives

Key findings (high-level)

Physics-consistency loss consistently improves physical accuracy (e.g., lower PDE residual and energy spectrum error) and can accelerate diffusion convergence.
Physics features alone are not consistently beneficial and may destabilize training.
Combining both can improve accuracy but may introduce instability under limited compute.

Dataset & evaluation

Dataset: 2D Kolmogorov flow (HR 512×512 / LR 128×128), Reynolds numbers 1000/2000/3000, forcing wavenumbers 8/12/16.

Metrics:

MSE (pixel reconstruction)
Physics Consistency Error (PCE)
Energy Spectrum Error (ESE)

Reproducibility

Training configs and scripts are included in the repository.
Visualization scripts compare HR/LR/reconstruction and energy spectra.

Structure-Based Drug Design via Diffusion Models Guided by Non-Differentiable Metrics

Thu, 01 May 2025 00:00:00 +0000

Method Overview

Zero-order Optimization via Gradient Approximation (SPSA)

Adaptive Trajectory Pruning (ATP)

Guidance with Spherical Gaussian Constraint for Conditional Diffusion

Mon, 01 Jan 2024 00:00:00 +0000

YesAI Lab, ShanghaiTech University

Thu, 01 Sep 2022 00:00:00 +0000

Overview

Research Member -ShanghaiTech University, 2022 – 2025*

Generative Modeling Research

Conducted research on diffusion-based generative models for image synthesis, image restoration, and controllable generation.
Participated in projects spanning computer vision and small-molecule generation under structured optimization objectives.
Explored model architecture design, training strategies, and generation-quality enhancement techniques.

Academic Research Workflow

Participated in the complete lifecycle of top-tier conference publications, including literature review, problem formulation, model design, experimentation, and manuscript preparation.
Co-authored a peer-reviewed publication as the third author.
Accumulated three years of continuous research experience in generative modeling and deep learning.

Experimental Development

Reproduced and extended recent research papers in diffusion models and generative learning.
Designed model improvements and conducted hyperparameter optimization for training stability and generation quality.
Performed systematic analysis of model convergence behavior and output fidelity.