Huachun Zhu Hirairi

• Level: L5.
• Received “Excellent”, the highest evaluation in a performance review, on Mar, 2023.

• Capstone project: Developing a full-stack editor powered by an AI copilot and AI content moderation system integrated with Langfuse in Troutwood Map at the CMU-founded fintech startup Troutwood.
• ML System course project: Composing KV Cache Compression Techniques for Long-Context LLM Inference.

• Selected in Su Buqing program: Dean's list at the School of Mathematics, and scholarships for outstanding students.

Nonlinear blend modes (e.g., overlay, multiply) enable rich visual effects but lack fast algorithms that decompose existing bitmaps into editable semi-transparent layers. We introduced a neural-network-based method that splits an image into linear and nonlinear alpha layers based on user-specified palettes and blend modes. Our method runs 370× faster than the state-of-the-art iterative optimization method while matching or surpassing its reconstruction quality.

We proposed the first method that enables latent‐space manipulation in generative adversarial networks under multidimensional conditions (e.g., keypoints, captions). Specifically, we designed an algorithm that searches for a new latent code that satisfies the target condition based on the Surrogate Gradient Field induced by an auxiliary mapping network. We also introduced a disentanglement metric for quantitative comparison, and demonstrated that our method achieved higher disentanglement across tasks including facial attribute editing and complex condition modalities.

We introduced a neural network based method for soft color segmentation, the decomposition of an input image into RGBA layers of homogeneous colors. Specifically, we trained a neural network to predict alpha masks and color residues in a single forward pass, using a novel objective that ensure proper color assignment amongst layers. Compared to iterative approaches, our method achieved up to 300,000× speed-ups with matching quality. We utilized our proposed method on several applications, and demonstrated its speed advantage, especially in video editing.

We introduced two GAN‐based editing operations that allow semantic edits over arbitrary image regions: (1) spatial conditional batch normalization with user‐specified weight maps and (2) feature‐blending on intermediate representations. Our approach works with any GAN with conditional normalization layers, can be used to edit both synthetic and real images, and is validated across multiple datasets and architectures.

We explored GAN‐based generation of anime faces by curating a clean, specialized dataset and applying DRAGAN regularization to stabilize training. Quantitative evaluations and case studies demonstrate high‐quality, diverse outputs. Our pretrained models are deployed with a web UI for interactive design.

• Led the development of Crypko (beta version), the world's first blockchain application of AI-generated Non-fungible tokens (NFTs).
• Developed the decentralized app with Vue.js frontend, Node.js backend, and Solidity smart contracts on Ethereum.
• Crypko was acquired by Preferred Networks in 2018.

• Trained CNNs, U-Nets, and GANs for classification, generation, and inpainting of line drawings using Chainer.

• Developed features in Digital Window, including smart scheduling, using Android/Java, UI in AngularJS, and backend in Ruby on Rails.

• Designed and trained arbitrage models on copper futures.
• Implemented, back-tested, and piloted trading algorithms in production.