Preparing research portfolio

FG / 26

Researcher · Computer Vision · Vision-Language Models · Generative Models

How machines see—and how we can maketheir vision more reliable.

Research trajectory

UnderstandEvaluateControl

Visual Reasoning

How can models ground and reason over objects and their relations in the real world?

Detected street entities connected as a scene graph — Grounded visual reasoning

Liu et al. Visual Instruction Tuning. NeurIPS 2023.

VLM Evaluation

How can we measure and calibrate the reliability of vision-language models?

Calibration curves and benchmark distribution for two multimodal models — Multimodal benchmark profile

Yue et al. MMMU: A Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark for Expert AGI. CVPR 2024.

Generative Models

How can we control generations while preserving fidelity and factuality?

A sequence progressing from noise to a resolved landscape — Rectified-flow generation path

Esser et al. Scaling Rectified Flow Transformers for High-Resolution Image Synthesis. ICML 2024.

01 / Selected projects

Projects, made visible.

Three project outcomes across scientific figure integrity, controllable generation, and diffusion-model analysis.

PROJECT / 012026

Pipeline for detecting semantic manipulation in structured academic figures

Project outcomeFengkai Gao

Detecting Semantic Manipulation in Structured Academic Figures: A Benchmark and a Structured-Extraction Baseline

Introduces SciFigSem-1K, a 1,000-sample benchmark for detecting contradictions between visual evidence and textual or numeric claims in structured academic figures, together with an auditable structured-extraction baseline.

Benchmark design · VLM evaluation · Structured visual reasoning

View project report View code

PROJECT / 022026

Detection-guided attention control method pipeline for text-to-image generation

Project outcomeFengkai Gao

Training-Free Concept Disentanglement in Text-to-Image Generation via Detection-Guided Attention Control

Proposes a training-free spatial-gated attention framework that uses open-vocabulary detection priors to reduce attribute leakage in text-to-image diffusion while preserving natural image composition.

Training-free control · Cross-attention · Concept disentanglement

View project report View code

PROJECT / 032025

Dynamic FreeU method pipeline with frequency analysis and feature modulation

Project outcomeFengkai Gao · Jiatong Zhu

Revisiting FreeU: Frequency-Domain Analysis and Dynamic Feature Modulation for Diffusion Models

Reproduces FreeU with precise feature intervention, validates its behavior through latent Fourier analysis, and proposes a time-aware Dynamic FreeU schedule that balances structural guidance with texture preservation.

Frequency analysis · U-Net features · Dynamic modulation

View project report

02 / Research directions

From visual evidence to controlled generation.

Computer Vision & Visual Reasoning

Recovering structure, relationships, and semantic consistency from complex visual evidence.

Explore field02

Vision-Language Model Evaluation

Designing controlled benchmarks that reveal how multimodal models reason—and where they fail.

Explore field03

Controllable Generative Models

Studying attention, feature modulation, and inference-time control in text-to-image diffusion.

Explore field

03 / Profile

Research grounded in how models actually behave.

I am an Artificial Intelligence undergraduate at Beijing Normal-Hong Kong Baptist University (BNBU). I research computer vision, multimodal reasoning, and generative models, and I enjoy turning model behavior into experiments that can be inspected and explained.

Emailt330034007@mail.bnbu.edu.cn

School

Beijing Normal-Hong Kong Baptist University (BNBU)

Major

BSc Artificial Intelligence

GitHubgithub.com/QCYTSN

Selected coursework

Machine Learning
Computer Vision
Natural Language Processing
Bayesian Networks
Big Data

Research directions

Computer Vision
Multimodal Reasoning
Generative Models