ruwwww

thoughts on software engineering, machine learning, and more

Inside ByteDance's Video Models: Building a Generative MLLM Planner from Scratch

ByteDance's video models (Seedance / Seaweed) are rumored to scale up to 200B parameters. Here is how their generative MLLM planning architecture actually works, and how we built a toy reproduction from scratch.

8 min read · August 15, 2026

2026 · generative-ai mllm diffusion-models computer-vision deep-learning · ML-Engineering Generative-AI
Training a Waifu Diffusion Model on Small, Corrupted Data using Patch Diffusion and Rectified Flow

How to train a data-efficient diffusion model on corrupted anime face data using CIELAB space, patch cropping, and modern transformer techniques.

12 min read · March 01, 2026

2026 · diffusion-models generative-ai flow-matching efficient-learning · ML-Engineering