We focus on Vision-Language Models, Diffusion & Generative Models, AI for Security and AI Safety, with a commitment to open-sourcing our research artifacts to advance the community.
📢 Migration Notice: This organization is the new home for all VELLDEPTH open-source projects (previously hosted under Alibaba-YuFeng). Please star/watch repos here for the latest updates.
- Diffusion & Visual Generation — acceleration, controllable generation, object removal, poster synthesis
- Multimodal Reasoning & MoE — routing analysis, cross-modal intervention, expert activation
- LLM Controllability & Steering — unified frameworks for parameter dynamics, behavioral evaluation
- AI Safety — multimodal adversarial robustness, safety benchmarking
- AI for Security — large language models for cybersecurity, domain-specific data synthesis and training
- Intelligent Vulnerability Analysis & Exploitation — automated vulnerability detection, analysis, and exploitation
- Intelligent Attack-Defense Adversarial — automated penetration testing, red-blue teaming, and adversarial simulation
| Project | Venue | Description |
|---|---|---|
| AttentiveEraser | AAAI 2025 Oral | Tuning-free object removal via self-attention redirection guidance |
| TC_Pade | CVPR 2026 | Trajectory-consistent Padé approximation for diffusion acceleration (up to 4× speedup) |
| DiffusionProbe | CVPR 2026 | Early-stage quality prediction for diffusion generation using CNN probes |
| SIMPLEPOSTER | CVPR 2026 | Minimalist inpainting framework for poster generation with faithful subject preservation |
| MMA-SafetyBench | — | Multimodal adversarial safety benchmark (under review) |
[2026.06] 🎉 Three papers accepted to ACL 2026 Main Conference:
- Seeing but Not Thinking: Routing Distraction in Multimodal Mixture-of-Experts
- Why Steering Works: Toward a Unified View of Language Model Parameter Dynamics
- How Controllable Are Large Language Models? A Unified Evaluation across Behavioral Granularities
[2026.04] 📄 We release the XekRung Technical Report — a frontier large language model for cybersecurity with comprehensive security capabilities, spanning continued pre-training (CPT), supervised fine-tuning (SFT), and reinforcement learning (RL). [Paper]
[2026.03] 🎉 Three papers accepted to CVPR 2026 Main Conference:
- SimplePoster — faithful subject preservation + position-controllable text rendering
- TC-Padé — trajectory-consistent rational approximation for diffusion acceleration
- Diffusion Probe — early-stage generation quality prediction
For questions or collaboration, feel free to open an issue in the relevant repository or reach out via email listed in our papers.
More open-source projects from VELLDEPTH are coming soon — stay tuned! ⭐