Kohaku BlueLeaf
Shih-Ying Yeh 葉適穎
CS master's student at National Tsing Hua University in Taiwan, also ML/NN engineer and core developer at Comfy Org. Published at ICLR 2024 (LyCORIS), ICLR 2026 (TIPO, ImagenWorld), ECCV 2026 (ReactVAU) and NeurIPS 2026 (TT-VidT).
I research and build AI systems end-to-end: from neural network training and GPU kernels to infrastructure, databases and agent systems. I also design compiler/IR/programming language for different projects, and design my own AI accelerator chip on my own accelerator framework (KohakuAccel/KohakuTPU). Open-source projects got over 8k stars on GitHub and 6.8M downloads on PyPI, models got over 3.8M downloads on HuggingFace in total.
Leading Kohaku Lab, a virtual open-source research lab. Currently working on agent framework (KohakuTerrarium), FPGA AI accelerator (KohakuTPU), video representation learning (TT-VidT) and infra for complex graph workflow with parallel/stream execution (KohakuNodeIR/Togeari). Everything we build is 100% open access.
"AI art should look like AI, not humans."
Recent Projects
LyCORIS
Lora beYond Conventional methods. Comprehensive PEFT library for neural networks, published at ICLR 2024.
KohakuTerrarium
General purpose AI agent framework with built-in tools, sub-agents, persistent sessions, TUI and web UI. Comes with kt CLI and package marketplace.
KohakuTPU [WIP]
AI accelerator implementation with Xilinx FPGA. Built on KohakuAccel, an open platform for building FPGA accelerators.
KohakuGBC
Togeari based implementation of Graph-Based Captioning (GBC). Builds graph annotation of each image (entities, regions and their relations) with VLM, OWLv2/SAM 3 and EmbeddingGemma, on one GPU, many GPUs or other nodes.
Togeari
Out-of-order dataflow machine for KIR graph programs (Turing complete). Runs thousands of KIR jobs at once on pluggable engines, batches calls across jobs to keep every engine busy, checks recorded runs against the reference interpreter.
TT-VidT
Decoupling the Temporal Axis for Efficient Motion-Centric Video Pretraining. Self-supervised video encoder for motion, published at NeurIPS 2026.
Recent Publications
TT-VidT: Decoupling the Temporal Axis for Efficient Motion-Centric Video Pretraining
Trinity: One Differentiable Physics for Training, Refining and Scoring Generative Floorplanners
World Editing: Intervening on Executable Worlds at Increasing Depth
Recent Career
ML/NN Engineer (Core Dev)
@ Comfy OrgCore developer of ComfyUI, tightly collaborated with comfy team to improve the usability and performance of Comfy UI eco system. The whole weight adapter system (LoRA/LoHa/LoKr/DoRA/OFT/BOFT) is created and maintained by me. Also working on lora trainer nodes, dataset processing nodes and FP4/FP8/FP16 training support.
Open Source Contributor
@ Kohaku LabDeveloped and maintained open-source projects related to computer vision and AI. Included personal project such as LyCORIS, TIPO, PixelOE, HakuLatent and more, and community projects such as stable-diffusion-webui (core dev), sd-scripts, ComfyUI (core dev) and more. Founded Kohaku Lab in Apr 2025, a virtual open source research lab for KohakuTerrarium, KohakuRAG, KohakuHub, KohakuRiver and more.
ML Engineer (Part-time)
@ AppleCollaborated with research team in Apple Inc. to explore the possibility of new form dataset. The work "Graph-Based Captioning: Enhancing Visual Descriptions by Interconnecting Region Captions" has been published on Arxiv and Apple official github.
