minWM

Diagnoses training/inference failures in video world model pipelines. Triggers on: training hang, loss NaN, shape mismatch, NCCL timeout, SP bug, distillation collapse, half-video noise. Use when user describes a symptom and wants root-cause diagnosis.

Step-by-step recipe for plugging a new video DiT backbone into minWM, grounded in the HunyuanVideo (HY15) and Wan 2.1 reference integrations. Use when a user wants to add a new backbone to the framework.

Onboarding guide for newcomers to video world model training in minWM. Covers two parts: Foundations (background theory for the two-phase pipeline) and Pitfalls (non-obvious mistakes from hands-on experience). Use when a newcomer is starting on controllable video generation training and needs background or wants to avoid common mistakes.