Publications | Yufei Liu's Homepages

SynPo: Boosting Training-Free Few-Shot Medical Segmentation via High-Quality Negative Prompts

Tue, 01 Sep 2026 00:00:00 +0000

SynPo addresses weak negative prompts in training-free few-shot medical image segmentation. It combines SAM and DINOv2 features to construct a synergistic confidence map, then selects informative positive and negative points through confidence-aware sampling and clustering.

Method	CHAOS-MRI					Synapse-CT
Method	Spleen	Liver	LK	RK	Mean	Spleen	Liver	LK	RK	Mean
SSL-ALPNettrained	67.02	73.05	73.63	78.39	73.02	60.25	73.65	63.34	54.82	63.02
ADNettrained	75.92	80.81	75.28	83.28	78.82	63.48	77.24	72.13	79.06	72.97
Q-Nettrained	75.99	81.74	78.36	87.98	81.02	74.86	71.21	75.26	74.79	74.03
RPTtrained	76.37	82.86	80.72	89.82	82.44	79.13	82.57	77.05	72.58	77.83
GMRDtrained	76.09	81.42	83.96	90.12	82.90	78.31	79.60	81.70	74.46	78.52
PerSAMtraining-free	69.14	42.44	64.84	71.36	61.12	65.03	65.55	58.47	60.31	62.34
ProtoSAMtraining-free	76.51	81.94	71.46	81.43	77.83	65.50	87.84	69.44	71.04	73.45
SynPo (ours)training-free	80.30	77.32	77.32	83.04	79.50	83.76	81.32	75.00	79.63	79.91

Qualitative results on abdominal MRI. Colored overlays show the predicted organ masks.

Module ablation

SAM	DINO	PSM	NRM	Dice	Gain
✓				57.48	–
✓	✓			66.70	+9.22
✓	✓	✓		78.74	+12.04
✓	✓	✓	✓	79.93	+1.19

Negative-prompt strategy on CHAOS

Model	Dice
PerSAM	56.72
+ Negative Point Selection	68.90

SynPo achieves a mean Dice score of 79.91 on Synapse-CT without task-specific training. The module study shows the largest gain comes from the Point Selection Module.

OmniPersona: Holistic Identity-Preserving Video Creation and Editing via Spatiotemporal Decoupled Persona Injection in Unified DiT

Wed, 01 Jul 2026 00:00:00 +0000

Role: Co-first Author (Equal Contribution)

Overview

Generation Example

The static input conditions and the animated output are combined into one synchronized video, following the presentation layout.

OmniPersona introduces Multimodal Holistic Persona Representation and spatiotemporally decoupled persona injection for unified video creation and editing. I co-led the work and contributed to the holistic identity representation module, combining face identity features with DINOv2 and CLIP features for clothing, hairstyle, and body characteristics.

On reference-to-video generation, OmniPersona reaches 79.3% face similarity and 68.7% body similarity, improving over the strongest baseline by 5.2 and 34.2 percentage points, respectively.

OneFont: A Unified Agent for End-to-End Font Creation

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

From Manual Model Selection to OneFont

OneFont completes model selection and refinement in a single agent workflow.

Motivation

Existing diffusion and MLLM-based font generation systems often focus on a single editing task. Supporting a new script or generation workflow can require substantial task-specific fine-tuning, while selecting the right model and repairing an imperfect result still depend on expert intervention.

OneFont treats end-to-end font creation as an MLLM agent task. Given a user request, it selects suitable models and tools, generates candidate results, verifies readability and style, and performs local refinement when needed.

Method Overview

OneFont converts each request into a reasoning trace and structured tool calls. Its generation library covers text-to-art fonts, style transfer, vector font generation, retrieval, handwriting generation, and local editing. Supervised fine-tuning teaches the agent to plan and invoke these tools, while GRPO-based preference alignment improves the quality of its generation strategy. During inference, a graph-based planner verifies intermediate results, backtracks when necessary, and applies local refinement.

Training framework and graph-based inference planner

Tool LibraryText, image, and answer tools for multimodal font creation

Model TrainingSFT builds tool-use ability, then GRPO aligns generation preferences

Model InferenceThe planner verifies results, backtracks when needed, and refines the output

Qualitative Results

Across text-rich generation tasks, OneFont produces readable text while matching the requested object, composition, and visual style. The comparison includes general image generators and dedicated text rendering models.

Qualitative comparison on cakes, T-shirts, book covers, and character images with embedded text. The final column shows OneFont results.

Clinical Study on the Application of a High-Sensitivity Electronic Nose on Thin-Film Gas Sensor Array Technology Combined with Deep Learning Algorithm for Early Non-Invasive Diagnosis of Chronic Atrophic Gastritis

Mon, 01 Sep 2025 00:00:00 +0000

This study evaluates breath analysis with a thin-film gas sensor array and learning-based models for early, non-invasive diagnosis of chronic atrophic gastritis.

TEmory: A Temporal-Memory Approach to Weakly Supervised Colonic Polyp Frame Detection

Fri, 01 Nov 2024 00:00:00 +0000

TEmory models relationships between adjacent video frames and stores representative normal and polyp features in parallel memory banks, improving robustness to minority normal structures in colonoscopy videos.

An Interpretable Approach to the Solutions of High-Dimensional Partial Differential Equations

Thu, 01 Feb 2024 00:00:00 +0000

The method transfers the structure of one-dimensional analytical solutions to higher-dimensional forms, then uses genetic programming and automatic differentiation to search for mathematically interpretable solutions that satisfy the governing equations and boundary conditions.