🎓 Second-year PhD Student at HKUST (Guangzhou), Robotics and Autonomous Systems (ROAS)
🔬 Research Interests: Generative Models, Reward Modeling, Vision Agents, Visual Perception
🧠 Current Focus: High-quality post-training, human-aligned reward modeling, multimodal agentic RL, and unified reasoning for visual understanding and generation
🌐 Website: ephemeral182.github.io
📧 Email: ephemeral182@gmail.com
I’m Sixiang Chen (陈思翔), advised by Prof. Lei Zhu and Prof. Fugee Tsung.
My work explores how generative and unified models can unlock new potential in image understanding and creation.
I believe research should matter to real users.
Some projects I lead — especially the PosterX series such as PosterCraft, PosterOmni, and PosterReward — aim to build dependable tools for designers and broader creative communities, while also pushing forward general post-training and alignment for image-text foundation models.
-
🧠 High-Quality Post-Training for Generative Models
Building post-training and alignment methods for high-quality image-text generation and editing, with a focus on reward modeling, preference learning. -
📏 Reward Modeling and Preference Alignment
Designing reward models that align AI outputs with human standards in aesthetics, structure, typography, and usability. -
🤖 Vision Agents & Unified Reasoning
Exploring multimodal agentic RL and unified models for open-world visual understanding, generation, and reasoning. -
🧩 Agent Systems & Harness Engineering Studying tool-using coding agents, skill/plugin ecosystems, memory, evaluation, and reliable orchestration through reproducible implementations.
🎯 PosterOmni · CVPR 2026
A unified framework for generalized multi-task poster creation and editing, covering both local refinement and global design.
🧠 PosterReward · CVPR 2026
A reward model for accurate evaluation and alignment of high-quality graphic design generation.
🎨 PosterCraft · ICLR 2026
Rethinking high-quality aesthetic poster generation in a unified framework.
Self-evolving image generation agents through tool-orchestrated visual experience distillation.
An evidence-aware daily Agent research radar with Chinese paper briefings powered by DeepSeek.
| Track | Repositories | Focus |
|---|---|---|
| Agent research | GenEvolve · GenEvolve-Training · Frontier Signal | Self-evolving agents, multimodal RL, paper intelligence |
| Agent harness lab | DeepSeek Harness · Pi · OpenHarness · Learn Claude Code | Agent loops, tools, plugins, orchestration |
| Skills & plugins | ASu Resume Skills · Pi Skills · Awesome Harness Engineering | Reusable skills, plugin patterns, learning references |
| Generative vision | PosterCraft · MeSa-IR · SnowFormer | Poster generation, image restoration, evaluation |
Forks in the Agent harness lab are curated as a study shelf. Original research and maintained projects remain the primary portfolio.
- Core loop — read Pi and Learn Claude Code for compact agent-loop implementations.
- Plugin architecture — compare DeepSeek Harness, OpenHarness, and Oh My Pi.
- Skills ecosystem — inspect Pi Skills, Awesome Pi Agent, and Awesome DeepSeek Agent.
- Harness engineering — use Awesome Harness Engineering as the index for memory, evaluation, permissions, observability, and orchestration.
- Feb 2026 — PosterOmni accepted by CVPR 2026
- Feb 2026 — PosterReward accepted by CVPR 2026
- Jan 2026 — PosterCraft accepted by ICLR 2026
- Jun 2025 — GenHaze accepted by ICCV 2025
- Apr 2025 — Released the GPT-4o Image Generation Capabilities evaluation report
- 2025 — Multiple papers accepted by CVPR / ICCV / AAAI
- 2024 — Papers accepted by NeurIPS / ECCV / CVPR, including CVPR 2024 Highlight
High-Quality Post-Training · Reward Modeling · Vision Agents · Agent Harness · Skills & Plugins · Image-Text Generation
| Type | Tools |
|---|---|
| Writing / Academic |    |
| AI Research Assistants |     |
| Development / Coding |     |
| Training / Infra |     |
| Design / Prototyping |      |
