Ruize Xia · Artificial Minds, Human Values

Text2Sign

Tue, 14 Jul 2026 00:00:00 +0000

Text2Sign is the public implementation behind the IEEE Access article of the same name. The repository provides a PyTorch training and inference path for short sign-language clips generated from text, using a frozen CLIP text encoder, a 3D backbone, factorized spatiotemporal attention, and DDIM sampling.

The design target is a single NVIDIA L4 GPU rather than a multi-node cluster. Evaluation uses a signer-disjoint How2Sign split so that appearance memorization is harder to confuse with text-conditioned motion.

Paper: and
Code:
Checkpoint:

CLIP Attention Analysis

Mon, 06 Apr 2026 00:00:00 +0000

This repository stores the experiment code for Matched-Learning-Rate Analysis of Attention Drift and Transfer Retention in Fine-Tuned CLIP. The controlled matrix covers CLIP ViT-B/32 on EuroSAT and Oxford-IIIT Pets, four shared learning rates, and five seeds — 80 completed runs in total.

The code records attention-drift metrics, in-domain accuracy, and adapter-aware CIFAR-100 zero-shot accuracy, then rebuilds the manuscript tables from JSON histories. CUDA and Apple MPS backends are both supported.

Paper:
Code:

On-Device Diffusion Kernels

Sat, 21 Mar 2026 00:00:00 +0000

This project turns Modulated Diffusion (MoDiff) from an operation-count story into a hardware result. The manuscript reports fused low-bit kernels and a cache-update fusion strategy that cuts extra memory traffic during iterative denoising.

On the evaluated setup, the implementation reaches up to 1.8× runtime speedup over FP32 and up to 42.2% lower memory I/O. The public tree lives at , forked from the official ICML 2025 MoDiff codebase and used as the systems implementation path for the paper.

Paper:
Code:

University Ranking

Thu, 22 Jan 2026 00:00:00 +0000

University Ranking is a product-shaped project rather than a paper: a searchable catalog of universities with bilingual names, theme and language preferences, and ranking data drawn from multiple public sources.

The backend is a REST API over SQLite, with ETL scripts that normalize institution names across datasets and expose filtering by name, country, city, and ranking scheme. The frontend is a React Native / web client with navigation, search, detail pages, and React Query for background refresh.

I wrote about the architecture and the “vibe coding” process on .

Frontend:
Backend: